What’s new

Gradients buys back and destroys 268 TAO worth of its own token

Gradients runs tournaments where people compete to train AI models, and entrants pay fees. Over the weekend the team used those fees to buy 268 TAO worth of Gradients' own subnet token, known as its alpha, and then burned it, meaning it was destroyed for good and taken out of supply. At this week's TAO price that comes to roughly seventy thousand dollars. The team says it does this regularly, and it is one of the few cases where money coming in from real users can be traced straight to a buyback.

Valeo trains a self-driving world model on NATIX driving data

Valeo, one of the world's largest car parts suppliers, trained an AI model called VATIX from scratch using driving footage from NATIX, the team behind StreetVision on subnet 72. VATIX is a world model: an AI that learns how a road scene is likely to unfold, which is used to train and test self-driving systems. Valeo trained more than 200 versions to work out how data, model size and computing power each affect the result. It then used those rules to predict how a much larger model would perform before building it, and the prediction came within 3.6 percent of the real result. The posts do not say how much of the data came through the subnet itself rather than NATIX's wider camera network.

Leadpoet opens a contest to build the best AI sales agent

Leadpoet, which uses AI to find sales leads for businesses, has launched its Lab Arena. Developers compete to improve how an AI agent picks out the right companies, spots signs that they are ready to buy, and backs each recommendation with evidence. Each winning agent becomes the new standard, its builder is paid, and its code is made public so the next competitor starts from a higher floor. The team also says LinkedIn and email connections are now built into the product, so users will be able to contact leads without leaving it, and that it is signing up new customers across several industries.

Hippius undercuts Google Drive and Dropbox on cloud storage prices

Hippius, a storage network where files are spread across many independent machines instead of one company's data centre, has launched new plans for its Hippius Drive app. Two terabytes costs seven dollars a month, against about ten dollars for Google Drive and ten to twelve for Dropbox, which is roughly 30 to 42 percent cheaper. Files are encrypted on your own device before they leave it, so the machines storing them cannot read them. A free plan gives ten gigabytes to try it. This is a product ordinary people can sign up for today, with a price anyone can compare.

NIOME opens a small test for swapping GENOME tokens into its subnet

GENOME is a token that trades on Base, a separate blockchain, and the project is moving its holders over to NIOME, the token of subnet 55. Instead of moving everyone at once, it has opened a small trial of the swap to prove the process works first. The trial is capped at 5 percent of GENOME's value at the snapshot, about 27,000 dollars in total. It is first come, first served, with a limit of 100,000 GENOME per wallet, and holders connect both their Base wallet and their Bittensor wallet to take part.

New tokens on Base and Robinhood's chain can now launch paired with TAO

Over three days the team at Forever Money connected TAO to several token launch tools. On Bankr, a tool that lets people create a token just by posting on X, new tokens on Base, Coinbase's blockchain, and on Robinhood's blockchain can now launch paired with TAO. That means they trade directly against TAO, so buying one means going through TAO first. A second launch site, BaseStonk, added the same option, and Forever Money's own website now lists every TAO-paired token. The team says each launch brings new TAO holders. That depends on whether these tokens draw real, lasting trading, and many of them will be meme tokens.

RedTeam says it is running five new trials with businesses

RedTeam turns miners into a round-the-clock network of ethical hackers, people paid to find weaknesses before criminals do, with a focus on defending against automated bot attacks. On the Bittensor foundation's weekly Novelty Search show, the team was billed as presenting five new trials with businesses and an upcoming product called Immune System. The businesses were not named, so it is hard to judge how big these trials are or whether any will turn into paying contracts.

Bitsec's tools find a serious security hole in an open source project

Bitsec, which built AI tools to check crypto smart contracts for security flaws, says the same tools found a critical flaw in an ordinary open source software project. The flaw would have let an attacker with basic access run any command on the machine, which amounts to a full takeover. It was reported and has been fixed. The project was not named. The point for Bitsec is that its tools can now look beyond crypto code to software in general.

Tatsu lets users describe a device model in a sentence instead of code

Tatsu builds AI models small enough to run on everyday devices. It has launched a Model Assistant, a chat tool where you describe in plain English what you want a device to do, and it either picks a ready-made model from Tatsu's catalogue or designs a new one that can then be trained. The aim is to let people without programming skills build these models. Picking from the catalogue is live now.

Enigma offers $250,000 to anyone who can break RSA-1024

qBitTensor, the team behind subnet 63 Enigma, has put up a $250,000 prize for the first person to break RSA-1024. RSA is the decades-old maths that scrambles data so only the intended recipient can read it, and the 1024 refers to the size of the key: it is an older, weaker setting that most banks and websites moved away from years ago, but no one has publicly cracked it. Enigma's work so far has been on solving problems that already have known answers while using very little computing power, so this is a step up into a problem nobody has solved in the open. Worth noting that the prize is an incentive, not a result: nothing has been broken. What we would watch is whether serious cryptography researchers actually show up to compete, since a bounty only tells you a team is confident enough to stake money on the attempt.

TPN Labs opens a contest to squeeze a large AI model onto an ordinary laptop

TPN Labs has started its third competition on Bittensor, called TPN-003. The task set for the miners, meaning the independent teams who compete for rewards on the network, is to take Gemma4-12b-uncensored, a large open AI model, and shrink it so it runs in under 8GB of memory while still scoring well across five standard tests. That 8GB ceiling is the point: it is roughly what a normal laptop or a consumer graphics card has, so a model that fits under it can run on a machine someone already owns instead of on rented data centre hardware. It is a contest rather than a finished product, so nothing is proven yet, and the thing worth watching is whether the winning entries hold up on the benchmarks or only look good on the scoreboard.

Numinous only pays forecasters who beat its own benchmark

Numinous, a subnet where people compete to predict future events, has changed how it decides who gets paid. It has added a baseline forecaster, meaning an in-house reference prediction that any competitor has to beat before they earn anything, alongside the price already quoted by prediction markets. Anyone scoring below that baseline now earns nothing, and the baseline itself will be raised over time by copying whatever the historically best competitors are doing, so the bar keeps moving up. The existing quadratic scoring, which pays out in proportion to how confidently right someone was, still decides how the rewards are split among those who clear the bar. The stated effect is that rewards concentrate on fewer, genuinely better forecasters. It could mean weaker participants stop earning and leave, which raises quality but shrinks the pool of people taking part.

SOMA plugs into GitHub Copilot and trims the text sent to the model

SOMA, subnet 114, says its early access is now open and its software sits behind GitHub Copilot, the coding assistant many programmers already use. When a developer asks Copilot a question, the assistant normally hands the AI model a large chunk of surrounding code so it understands the context. SOMA squeezes that chunk down before it is sent on to the DeepSeek V4 Pro model, and claims this cuts the amount of text going in by roughly a tenth. Since AI services charge by the amount of text they process, less text in means a smaller bill for the same answer, and the developer sees no change to how they work. The claim of about ten percent is SOMA's own and has not been checked independently, so it is worth watching whether users report the same saving once early access widens.

Minos makes its scoring checkable by anyone

Minos, the subnet that scores how well its participants perform, has published the machinery behind that scoring so outsiders can check it. Three things changed. Each scoring run now publishes a config hash, a short fingerprint of the exact settings used, so nobody can quietly change the rules afterwards and claim they were always that way. The tasks that get scored are picked on the blockchain itself from a window of options announced in advance, rather than chosen behind closed doors. And after each round the team releases the truth files, the correct answers the work was marked against, so anyone can rerun the marking and see whether they get the same result. This came after criticism in the project's own chat about whether the scores could be trusted, and the response was code rather than argument. Marking is where a subnet decides who gets paid, so being able to audit it is the difference between taking the team's word and checking for yourself. Worth noting that the accounts reporting this are community members rather than the Minos team, so the claim is theirs, not the project's.

SayGm adds Google's new Gemini 3.8 Flash model at a small discount

SayGm, the subnet that resells access to big AI models, says Google's newly released Gemini 3.8 Flash is now available through its service at 8.3% below Google's own list price, and that it has also switched on image generation for Gemini models through its API, the connection other people's software uses to send requests automatically. This is the routine work of a reseller: match a new model on the day it appears and undercut the source slightly. It is worth noting only because same-day availability is the thing this subnet is competing on, and a small discount is the whole of its pitch to developers deciding where to send their traffic. There is no revenue figure or customer named here, so it says nothing yet about how many people are actually buying.

Ditto says it is paying for Hippius storage and Desearch search in production today

Ditto, a company building on Bittensor, says it is not announcing a partnership but simply paying two subnets for work it already relies on: it stores its data on Hippius, a decentralised storage network where independent operators host files, and it routes its search queries through Desearch first. Both teams stressed the point that this is ordinary production usage with real money changing hands, not a press release. That distinction matters because most subnet news is about intent rather than income, and a customer who pays every month is the clearest evidence a subnet has something people actually want. Hippius says a fuller explanation of how Ditto's storage runs on its network is coming. What we would watch is whether the spend is disclosed in numbers, since without a figure it is impossible to tell a serious contract from a small test.

Bitcast spent $75,000 buying back its own token in August

Bitcast, the subnet that pays creators to promote brands on YouTube and X, says it used $75,000 of the money it earned from customers to buy its own token back off the open market in August. That is its biggest month so far and more than double what it spent in July. A buyback means the team takes real revenue and uses it to purchase the token that its network runs on, which takes some supply out of circulation and ties the token more directly to how much business the subnet is actually doing. The figure comes from Bitcast itself rather than an outside auditor, so it is worth watching whether the monthly numbers keep climbing or whether August was a one-off peak.

True Performance Network turns on its public scoreboard and gateway

Subnet 65, run by TPN Labs, says three pieces are now live two weeks after it started running on the main network: TPN Bench, a public dashboard, and the TPN Gateway. The subnet pays people to shrink large AI models so they run cheaper and faster, and 17 of those participants have squeezed a model called Qwen3.5-2B down to a 2GB limit. Until now the judging of who did that best happened out of sight. A public dashboard and a repeatable benchmark mean anyone can check the scores rather than take the subnet's word for them, and the gateway is the door through which outside software can actually call the compressed models. It is a plumbing release rather than a headline one, but verifiable scoring is what turns a contest into something a customer can buy from.

Yanez says its private sign-in is now live with two partner apps

Yanez, the subnet that works on proving who someone is without collecting their personal details, says its YID login has begun rolling out inside two other projects: dFusionAI Personas and Syklo's two-factor security check, which is the extra step some apps ask for on top of a password. Yanez reports that in the first stages of the rollout only 2% of people signing up needed to contact support, which it offers as evidence the process is not confusing people. The pitch is verification where no identity documents change hands and there is no password to lose. Two live integrations is a small sample and the numbers come from Yanez itself rather than an outside check, so the thing worth watching is whether the same low support rate holds once far more people are put through it.

Zeus says the selling of its token was one backer leaving, and it is finished

Zeus told holders that the recent run of selling in its token came from a single early partner who had paid the fee for Zeus to get its slot on the network, and in exchange received a share of the newly created tokens the subnet earns each day. That partner has decided to wind down and move on, and Zeus says the selling connected to it is now over. Zeus said it does not agree with the decision but respects it. This is the kind of detail a reader normally never gets: an explanation of who was selling and why, rather than silence. It does not tell you anything about what the token does next, only that this particular source of supply has stopped.

Tao.com says its new automatic sell orders are its most used feature this quarter

Tao.com, a trading site for TAO and subnet tokens, says the take profit and stop loss tools it added recently have been the most requested and most adopted feature it has launched this quarter, and that its best performing traders are using them. Take profit and stop loss are standing instructions to sell automatically if a price rises to a chosen level or falls to one, so a person does not have to sit and watch. The company says more advanced tools are coming. This is a product update on one venue rather than a change to Bittensor itself, and the claim about who is using it comes from the company with no figures attached.

Open Quantum rewrites its whitepaper after nine months of changes

The team behind Open Quantum, the quantum computing subnet, has published a substantially expanded version of its whitepaper, the document that explains what a project is trying to do and how it works. They say the platform has changed a great deal since it launched in December and the old paper no longer described it. A rewritten document is housekeeping rather than news in itself, but it is the main thing an outsider can read to judge whether a technical project is doing what it claims, and letting it go stale for most of a year is common in this space.

Enigma opens up to people who do not want to run software

Enigma, subnet 63, works on quantum computing problems and until now taking part meant running miner software, a program that has to stay connected to the network and needs a crypto wallet configured before it will do anything. That is a real barrier for a researcher who has an answer but no interest in the plumbing. The team says submissions can now be made through a web page instead: upload a zip file, click a few times, and validation starts. The claim worth watching is whether the people it brings in produce better work than the miners already there, because lowering the entry cost also lowers the cost of submitting noise. For now it is a straightforward widening of who can compete on one subnet.

Cacheon reports a new speed record in its own model contest

Subnet 29's Cacheon runs what it calls an arena, a competition where entrants try to make an AI model run faster without breaking it. The team says the latest round reached 2,274 tokens per second on a MiniMax-M3 baseline. Tokens are the small chunks of text a model produces, so tokens per second is simply how fast it writes, and the baseline is the unmodified starting version everyone is trying to beat. Cacheon says it judges entries only by end to end testing, meaning the whole system has to work rather than one part scoring well in isolation, and that it is now adding the next model to the arena. This is a routine progress update on one subnet's own benchmark, not an outside result, so the number is best read as their own scoreboard moving.

The Claims subnet goes live and starts earning

Subnet 111, run by DeSci Claims, says its main network launch is done and that TAO emissions to it have been switched on. Emissions are the new TAO the network creates and hands out every few seconds, split between the people doing the work on each subnet and the people who back them, so the switch being flipped is the moment a subnet stops being a test and starts being paid. The team has not yet published numbers on what the subnet is producing, so the useful thing to watch from here is whether real output follows the money rather than the launch itself.

Kusanagi sets up as a business layer around Bittensor rather than a subnet

A group called Kusanagi has launched itself as a growth layer around Bittensor, meaning it is not registering a subnet of its own but bundling four jobs under one name. It runs a validator, one of the machines that scores work and votes on who gets paid, rebranding the existing MUV validator as its own. It keeps TaoSwap, a site for buying and browsing subnet tokens, as its shop window. It offers to work with subnet teams on strategy and marketing, which is a service most subnets currently do without. And it says it is aiming at a regulated investment vehicle so that traditional money can hold TAO and subnet tokens without anyone having to set up wallets and learn the tooling first. Only the first three exist today, and the regulated vehicle comes with no named regulator, jurisdiction or date, so it is worth treating that part as an ambition until someone else confirms it.

Babelbit is dropped from the network a week after launching its dubbing tool

Subnet 59, Babelbit, has been deregistered, meaning it lost its slot on the Bittensor network and no longer receives any of the newly issued TAO that the network pays out to the projects running on it. It happened roughly a week after the team announced a dubbing service for translating audio into other languages and claimed progress on a model that handles many languages at once. Losing a slot works a little like losing your pitch at a market: the stall has to outbid others to stay, and Babelbit did not. It is a reminder that announcing a product is not the same as holding a place on the network, and that a subnet can disappear while its news is still fresh. This comes from an outside tracking account rather than the team itself, so their own explanation has not been heard yet.

Targon gave the new Cascade subnet free computing to get started

Targon, which rents out computing power on Bittensor, says it handed the team behind subnet 91, Cascade, free compute credits to help them launch. Cascade is the first team to come out of BitstarterAI's machine learning track, a kind of training programme for new subnet teams, and it is building forecasting models: software that learns from a long run of past numbers, such as prices or demand, and predicts what comes next. Nothing has been shipped for the public to use yet, and no money changed hands that we can see, so this is a starting line rather than a result. What it does show is that the ecosystem's own compute providers are subsidising new teams instead of charging them from day one, which lowers the cost of trying something on Bittensor.

Numinous buys back its own token with money earned from customers

Numinous says it has made its first buy-back, spending 4 TAO on its own alpha token, the subnet-specific token that tracks how the network values it. The money came from real income: the team says it is 10 percent of what its Forecasting API, a paid service other people call to get predictions, took in across July and August, and it published the transaction so anyone can check it. The sum is small in absolute terms, so this is not the size of the purchase mattering, it is the fact that a subnet earned outside revenue and put a fixed share of it back into its own token, and said it will keep doing so. What we would watch is whether the next few buy-backs arrive on schedule and grow, which is the only thing that turns a gesture into a policy.

Trishool is let back into the Arbos root basket after a burn was misread

Trishool says it has been added back into the Arbos root basket, one of the managed baskets that decide where pooled root stake gets pointed. It had been dropped earlier because it burns 80 percent of what it earns, meaning that share of its token issuance is destroyed rather than paid to anyone, and that pattern was read as a sign the subnet was rewarding the wrong behaviour. Trishool's account says it showed the evidence that the burn is genuine destruction rather than a payout in disguise, and that Arbos re-checked and reversed the decision. The wider point for readers is that these baskets are run by people making judgement calls on incomplete information, and a subnet can be scored out on a misreading and then scored back in.

TensorUSD's dollar token now appears in SubWallet without any setup

TensorUSD says its TUSDT token, a coin meant to hold a steady value of about one US dollar, now shows up automatically inside SubWallet, a popular wallet app for this family of networks. Previously, getting a token to appear often meant copying a long technical address and adding it by hand, a step that puts off newcomers and is easy to get wrong. Now a user switches the wallet to Bittensor and the balance sits next to their TAO, ready to send, swap or track. It is a convenience improvement rather than a new capability, but the friction it removes is exactly the friction that stops non-technical people using a token at all.

Bittensor writes down who actually produces its blocks, and says that should change within a year

A documentation page was merged into Bittensor's core code that, for the first time in one place, states plainly how the chain itself is run. Right now it uses Proof of Authority, meaning a known, permissioned set of machines takes turns producing blocks and confirming them, rather than anyone with enough stake being able to do it. The page also sets out a non-binding aim to move to Nominated Proof of Stake, where block producers are chosen by how much TAO is staked behind them, within roughly the next year. It carefully separates this from Yuma Consensus, which is the separate scoring system that decides how subnet rewards are split. Nothing about the network changed today: this is a description, not a switch being thrown. What it could mean is that a part of Bittensor most holders never think about, who is trusted to keep the chain running, is now openly acknowledged as centralised with a stated intention to open it up. Worth watching is whether that timeline turns into actual code, since a roadmap target that is explicitly non-binding is a statement of intent and nothing more.

Nodexo adds 22 top-end GPUs, with 14 free to rent immediately

Nodexo, the subnet that rents out data centre computing power, says it has added 22 H100 graphics cards to its network, and that 14 of them are sitting unused and available to hire right now with no waiting list or contract. An H100 is the expensive Nvidia chip that most serious AI training and large model work runs on, and access to them is the main bottleneck for anyone building AI outside the big cloud companies. The claim that came with it is that the capacity is genuinely idle rather than booked out, which is the useful detail: a rental network is only worth anything to a customer if there is spare stock on the day they need it. What it could mean is more paying work flowing through the subnet if that capacity gets taken up, and the thing to watch is the opposite reading, that hardware sitting free is hardware nobody is paying for yet.

Subnet 3 starts a 110 billion parameter training run on strangers' computers

Teutonic, which sits on subnet 3, has begun training a model called Teutonic-II with about 110 billion parameters, the internal settings a model adjusts as it learns and a rough proxy for its size. The run is permissionless, meaning anyone with suitable hardware can join in and contribute computing power rather than the work happening inside one company's data centre, and the team says it is roughly ten times the size of their last model and larger than any earlier decentralised run they are comparing against. They have also said they will publish checkpoints, part-trained snapshots of the model, while the run is still going, plus the final dataset it was trained on. The news came through the Opentensor Foundation's weekly roundup. What it could mean is a real test of whether a network of independent machines can hold together a job this big without a central operator, and the questions worth watching are whether the run finishes, how the finished model compares with ones trained the conventional way, and whether those promised checkpoints and data actually appear.

SayGm plugs SOMA's matching algorithm into its own app

SOMA, the subnet building software that pairs people or content with what suits them, says another Bittensor project, SayGm, has started using its algorithm inside its own product. Subnets are the independent teams that make up the Bittensor network, and they usually build alone, so one team's work being adopted by another is a small sign that the output is good enough for someone else to depend on. SOMA also says an early access version of its own product is coming soon, without giving a date, and is floating an idea it calls hybrid competition, where miners would be judged on more than one measure at once rather than a single score. The integration is the concrete part here: the product date and the competition change are still just intentions.

Ditto ships a search that returns decisions, not documents

Ditto, a Bittensor project, says it has shipped a new way of searching a team's own records. Most search tools look for where a word appears, so asking what your team decided about pricing in March returns a pile of documents that happen to mention pricing. Ditto's version combines three approaches at once: plain keyword search, embeddings, which match on meaning rather than exact wording, and a graph, which records how people, decisions and events connect to each other. The stated aim is to hand back the decision itself. It is a real shipped feature rather than a plan, though the claim about quality is the team's own and has not been independently tested, so the thing to watch is whether users report it actually finding the answer.

Upgrade 451 never made it to the live network, so its work now rides inside 452

Earlier this month the developers opened a branch for network upgrade 451, the numbered package of changes that every machine running Bittensor eventually installs. It was built but never got the two signatures it needed to go live on the main network, and two attempted release runs were cancelled. Rather than keep waiting on it, the team has merged a change that renumbers the current code to 452, folding 451's work plus a small internal fix into one package that can ship on its own. Nothing about the network has changed yet: 452 is still sitting in the queue waiting for signatures. What it means in practice is that 451 as a separate release is gone, and anyone tracking which upgrade is next should now be watching 452.

The tightening on built-in shortcuts is now merged

A change we reported earlier as a proposal has now been accepted into the main code. Bittensor lets Ethereum-style smart contracts reach network features through built-in shortcuts, called precompiles. The fix means a shortcut will now only run if it is genuinely being called at its own address, rather than being triggered from somewhere it does not belong. Simple cryptographic ones, which do not touch any stored data, are exempt because they cannot be misused this way. It brings Bittensor in line with the checks Ethereum tooling already applies. It is a defensive tightening rather than a response to anything going wrong, and it will reach the live network as part of upgrade 452.

Engy undercuts the official price of the new GLM-5.3 model by about a third

Engy, subnet 53, is now serving GLM-5.3, an AI model released this week with its weights published openly so anyone can download and run it. Engy says it is running at roughly 100 tokens a second, meaning about 75 words of output per second, and charges 98 cents per million words of input and $3.08 per million of output. The official service and the reseller OpenRouter charge $1.40 and $4.40 for the same amounts, so Engy is asking around 30 percent less. This is the second Bittensor provider to carry the model within days of its release. It is a concrete, checkable claim rather than a promise, and it is the kind of evidence that matters for the wider argument that a network of independent operators can undercut the company that made the model. What we would watch is whether the speed and the price hold once real demand arrives.

SayGm serves the new GLM-5.3 model on launch day

SayGm, subnet 28, says it is already running GLM-5.3, an open AI model whose files were released publicly the same day. The model is hosted through KubeTEE, a partner that runs it inside hardware designed so the operator cannot read what users send. This is the latest in a run of day-one launches from the subnet, and the practical claim is speed: an ordinary user can try a brand new model through SayGm within hours of it existing, rather than waiting for a large provider to add it. It is a normal product update rather than a change to how anything works.

Root Reborn claim fixes stop odd cases blocking ordinary payouts

A change to Bittensor's core code has been merged and is now part of the codebase. It tidies up three things in Root Reborn, the reworked system that decides how people who stake to Root get paid. First, it makes sure that a basket of holdings that has run dry or grown unusually large cannot jam the queue and stop healthy claims going through. Second, it adds tests that prove the accounts run by subnet owners keep their Root rewards rather than having them burned, which is worth checking because a mistake there would quietly cost those teams money. Third, it removes a leftover check for a migration that has already finished. None of this changes what anyone is owed, it is housekeeping that makes the payout path harder to trip up.

The shortcut for new machines joining the network is now merged

We reported this as a proposal, and it has now been merged into Bittensor's code, so the status has changed from proposed to done. When a new computer joins the network it normally has to work through the entire history of the chain before it can be useful. Warp sync is a shortcut that lets it jump most of the way by following a trail of signed handover records instead. A one-off historical quirk on Finney, the live Bittensor network, broke that trail and left the shortcut unusable. The fix writes a trusted, known-good checkpoint into the network's settings file so the shortcut has a solid place to start from. For anyone running or setting up a node it means far less waiting, and for everyone else it means fewer barriers to new machines helping to run the network.

Root fund managers can now choose their subnets, with a one sixteenth ceiling on any single bet

When Root Reborn launched, the people running these funds were not allowed to say which subnets their fund should back: the setting was switched off. Runtime 449, now merged, turns it back on but adds a guardrail. No fund can put more than one sixteenth of its weight, a little over six percent, into any single subnet. The point of Root Reborn was to stop everyone's stake piling into one place, and without a cap a fund manager could have rebuilt exactly that concentration by hand. The open questions are whether one sixteenth is loose enough for a manager to add real value, and whether curation attracts stake back to Root or simply gives fund runners a new lever worth competing over.

Emergency fix stops one stuck subnet blocking everyone's Root payouts

Root Reborn is the system that lets people stake to a fund which spreads their TAO across several subnets rather than picking one by hand. A hotfix, runtime 451, has been merged after developers found that if one subnet in a fund had no usable liquidity, meaning there was not enough of its token available to trade in and out, the whole claim could fail and healthy positions alongside it could not be settled either. The fix now treats a dead-end subnet as a known outcome and pays out the rest, and it adds a test proving that subnet owners do not lose their own Root rewards. It matters because a fault like this quietly stops people getting money they are owed, and the answer arrived as a same-day hotfix rather than waiting for the next scheduled upgrade.

Fund prices get one agreed way of being written down

Different tools have been showing the price of a Root fund in different units, which makes two dashboards disagree about the same number for no real reason. Release 450 adds a single on-chain reference point, recording for each fund the starting rate and the block it started from, so every wallet, explorer and bot can convert to the same figure. Nothing about how much anyone owns changes. It is plumbing, but plumbing that decides whether the number you read in one place matches the number you read in another.

New machines joining the network get a trusted shortcut to catch up

A computer joining Bittensor has to work through the chain's history before it can take part. There is a fast path for this, but it needs a known good starting point, and without one new nodes could stall trying to verify their way in. This merged change ships that starting point inside the network's configuration file, a historical checkpoint plus the signatures that back it, so the fast catch up works out of the box. It lowers the effort of running your own node, which matters because a network is only as decentralised as the number of people who can afford to run one.

Bitcast's numbers put revenue at 346 thousand dollars, with most recycled into its token

A widely shared breakdown of Bitcast, the subnet that pays video creators for making content about sponsors' products, puts hard numbers on its progress: 974 creators with a combined audience of about 11 million, 6,000 pieces of content across 114 campaigns, 346 thousand dollars of revenue, and 202 thousand of that used to buy the subnet's own token. We already covered the creator signups passing one thousand; what is new here is the revenue and buyback figures, which are the part that matters, since actual money coming in and being recycled into the token is what most subnets still lack.

Tatsu trains three small AI models on real sensor fleet data

Tatsu, which builds TatsuOS for connected devices, says it has built, trained and tested three AI models on real data from device fleets: one that spots anomalies, one that classifies what state a device is in, and one that forecasts what happens next over short horizons. The models are shrunk down to kilobytes so they can run on the devices themselves rather than in a data centre, and the team says the cloud infrastructure that trains and ships them is now stood up. It is an ordinary weekly progress update, but a concrete one: the product is moving from collecting sensor data to interpreting it.

Bittensor's core code now comes with a rulebook for AI coding assistants

The main repository behind Bittensor has merged a written set of instructions aimed at the AI agents that increasingly help write code for it. The reason given is that submitted changes kept failing the same automated checks over and over: code formatted the wrong way, generated files left out of date, dependency lists not locked to exact versions, documentation not rebuilt. The guidance requires those checks to be run before a change is put forward rather than after it fails. It is housekeeping rather than a change anyone holding TAO will see, but it is a quiet sign of how much of the protocol's own code is now being drafted with machine help, and of the maintainers deciding to set rules for it rather than leave it informal.

Bitcast says a thousand creators have signed up to earn on its network

Bitcast, the subnet that pays video and social creators for the audience they bring rather than for computing power, has reached a thousand registered creators according to a figure shared by the project's lead. That number matters more than it sounds, because most subnets are worked by people running specialist machines, and Bitcast's whole bet was that ordinary creators with no technical background could take part if the sign-up process was simple enough. A thousand registrations is proof that people are turning up, though it says nothing yet about how many are actually posting or how much they are earning, which is the number worth watching next.

A sync fault on the practice network has been fixed

Developers have merged a fix for the test network, the practice copy of Bittensor where changes are tried out before they touch real money. The fault was in the fast catch-up process a new machine uses to download the network's history without replaying every step from the beginning: it was checking one historical point against the wrong list of approvers, so the shortcut failed and the machine could not join. The fix points it at the correct list. This affects only people setting up test machines and changes nothing about the live network, but a broken test network slows down everything that gets tried out on it first.

Dropbox is trialling LeadPoet to find its sales leads

LeadPoet, the subnet at number 71, says Dropbox and one of its sales partners have started a pilot using its system to find sales leads. A subnet is an independent project that rents its work out through the Bittensor network, and LeadPoet's job is to hunt through public sources for companies showing signs of wanting to buy, then check that the contact names and job titles are still current. A pilot is a trial, not a signed long-term contract, so the thing to watch is whether it turns into paid ongoing work and whether Dropbox itself confirms it. What makes this notable either way is the customer: a well known software company using a Bittensor subnet for real day to day business rather than for an experiment.

Safety checks added to the kit developers use to move stake

The Bittensor team merged a set of fixes to its SDK, the ready made code that wallets, bots and tools use to talk to the network so they do not have to write everything from scratch. Six gaps were found while reviewing the upcoming version 448 software, all of them in places where a request could be sent in a state the network would not accept: private submissions that are hidden until they are processed, claiming staking rewards from root, and the sequence where someone claims rewards and then withdraws their stake in one go. The kit now refuses those requests up front instead of letting them fail somewhere later. Nothing changes for people who only use a wallet, but it lowers the chance of a tool built on top of Bittensor losing or stranding a transaction during the switch to the new version.

Green Compute says its GPU rental subnet is now live

Green Compute announced it has launched, and the post drawing attention to it lays out an unusual money flow: customers pay for long term access to GPUs, the specialist chips that AI models run on, and that income is used to buy the project's own token, called alpha. Miners on the subnet have bought the hardware themselves and take years to earn it back, which the supporter argues gives them less reason to sell what they are paid. The pitch is that demand comes from outside crypto, from businesses that want computing power rather than from traders. Two things are worth watching before treating that as proven: we have the team's own launch wording only second hand through a quoted post, and no customer, contract or revenue figure has been named, so the buying of the token described here is a design rather than something anyone can yet measure.

A version clash that was blocking the new software release has been cleared

The automated process that publishes Bittensor's Python software, the version most developers install, failed when it tried to ship the code for the coming version 448 upgrade. The cause was mundane: one of the packages still carried a version number that had already been published, and the system refuses to release the same number twice. The number has now been raised to 0.1.4 across the files that declare it, and the release can go out. Nothing about how the network behaves is affected, but it removes an obstacle to developers getting the tools that match the upgrade waiting to go live.

Moving stake between subnets can be capped at a price you choose

Staking means locking TAO behind a subnet so it earns a share of that subnet's rewards, and each subnet has its own token whose price moves as people buy and sell it. That means a large move of stake from one subnet to another can push the price against you while it is happening, and finish at a noticeably worse rate than the one you saw when you started. Code now merged into release 448 lets you name the worst price you are willing to accept and then pick one of two behaviours: take the whole amount at that price or none of it at all, or take as much as fits under your limit and leave the rest. This is standard protection in ordinary trading and its absence has been a real cost for anyone moving size. It is written and merged, but only takes effect once the 448 upgrade is signed and goes live.

One command now closes every staking position at once

Stake is usually spread out: across several subnets, and across several hotkeys, which are the named keys identifying who you have staked to. Getting out completely used to mean repeating the exit for each one by hand, which is slow and easy to get wrong. A merged change to the command line tool adds an option that finds every live position you hold, removes any duplicates, and submits them all together as a single batch so they either all go through or none do. It is a convenience most of the time, but it also matters when someone wants out of something quickly and every extra step costs them minutes.

The tidy-up that stops the chain remembering old stake links is now merged

Previously reported as a proposal, this change has now been merged into the code. The chain keeps a note of which staking keys a wallet is linked to, and those notes were left behind even after the stake itself was gone, so the records slowly filled up with connections that no longer meant anything. The merged version clears each link once the pair genuinely has nothing staked anywhere, while keeping it if stake still exists on another subnet. It is housekeeping rather than a change anyone will feel, but stale records are what make chain data hard to read and slow to search over time.

Teutonic releases a 10 billion parameter model trained across scattered machines

Subnet 3, Teutonic, has put out Teutonic-I, its first model trained from scratch rather than fine-tuned from someone else's. The interesting part is not the size but the method: it was trained across many independent machines contributed by strangers rather than inside one company's data centre, which is far harder to make work. On the benchmarks used to compare such models, it came first on 8 of 11 tests and beat models several times its size. Benchmarks are self-reported and chosen by the people reporting them, so the number to watch is whether outside developers pick it up and get the same results. If they do, it is evidence that pooled, permissionless training can produce models that stand up next to conventionally trained ones.

Subnet 88 says its fund and index are live and earning, and buying of its own token has started

Investing 88 says it has moved past the roadmap stage into products people can actually put money into: an 88 Quant Fund launched at the end of 2025 and a public index opened at the end of March, with the index capped at 10,000 TAO of money under management. The team also acknowledged that buying of its own subnet token has begun on-chain. Every subnet issues its own token, and a team using revenue to buy that token back is one of the few signals that money is genuinely coming in from outside rather than circulating within the network. Two caveats: the cap on the index means the amounts involved are still small, and none of the revenue figures behind the buying have been published in detail.

Babelbit says it can add new languages without building each pair by hand

Babelbit, which works on real-time translation, says it has reached a point where it no longer has to build and tune a separate system for every language pair. Until now, going from English to Japanese and English to Polish meant two separate builds. The team says it is moving to a single multilingual foundation that covers many languages at once, which is how the largest translation systems work and is the difference between adding a language in months and adding one in days. No independent test results have been published yet, so the claim rests on the team's own account. The measure to watch is how many languages actually appear, and how quickly.

ByteLeap's share of the network's payouts is switched back on

Subnet 128 says its emissions have been officially restored. Emissions are the daily TAO the network pays out to each subnet, split between the people running the machines and those backing it, so losing them means a subnet earns nothing while it stays registered. The team did not spell out why the payments stopped, but did admit it had failed to keep its community updated and promised to start disclosing revenue and business data from now on. For anyone holding that subnet's token, the practical change is that rewards flow again, and the promise of published numbers is the thing to hold them to.

Vidaio splits its video technology into five products aimed at different industries

Vidaio has explained how it is packaging what it builds. Rather than selling one general video tool, it is taking the same underlying stack of video processing and delivery technology and configuring it five different ways for five different markets, including a localisation product with its own dashboard and a compression product of its own making. Localisation means adapting video for other languages and regions, and compression means shrinking video files so they cost less to store and stream. The team says the localisation product gets its first full public showing at IBC, the large broadcast industry convention, in September. Packaging alone is not revenue, so the thing worth watching is whether any of the five come back from that convention with named customers.

Score's tools get built into other Bittensor projects, including a paying contract

The team behind Score says its Score Studio product now has working integrations with Hippius, a decentralised storage subnet, and gm, and that it has a working contract with the video subnet Vidaio. This matters more than it might sound. Most subnets sell to the outside world or to nobody, and one subnet actually paying another for a service is a sign that these projects can be useful to each other rather than only competing for the same pot of network payouts. The terms of the Vidaio contract have not been made public, so the size of it is unknown.

Bitstarter turns on a validator that funnels its fees back into new subnet teams

Bitstarter has switched on a validator, which is one of the machines that checks work on the Bittensor network and takes a cut of the payouts for doing so. The unusual part is what it says it will do with that cut: share earnings with the teams it backs and put the fees into buying those teams' tokens, rather than keeping them. The same group says its Pitchtensor competition raised over 800 TAO in its first round and produced two new subnets heading for launch, and it is running the format again at another conference aimed at machine learning teams. It is an attempt to build a funding pipeline for new projects on the network, and the test of it is whether the two subnets it has already funded reach launch and stay there.

Score Studio labels a raw sports clip automatically in a live demo

Score has published a video showing its Studio product taking an unedited volleyball clip and automatically generating the labels and outlines that a computer vision system needs in order to learn from footage. That labelling work is normally done by hand and can take weeks, and it is the expensive part of teaching software to recognise anything in video. In the demo the automatic result still needed correcting, with the two teams not cleanly separated, so this is a working tool rather than a finished one. It is a useful look at what the subnet actually produces, which is more than most subnets show.

Ditto replaces its scoring system after miners gamed the last one

Ditto has released version 11 of Dittobench, the test it uses to score the AI agents competing on its subnet. The team says version 10 was maximised by miners within hours of going live, meaning competitors found ways to score highly on the test without being better at the underlying task. Rewriting the test is the standard response, and the speed at which it had to happen is a fair illustration of how quickly people find the edges of any scoring rule when money is attached. The team says coding ability is the next skill it will add to the test.

Tatsu gets its software running on boards with no wireless of their own

Tatsu, which puts its software on small cheap computers, had until now only supported boards with wireless networking built in. It has added support for the Raspberry Pi Pico, a popular board costing a few pounds that has no radio, by pairing it with an inexpensive add-on wireless module and building a firmware profile around that combination. The point is not the one board: it means the platform can run on the much larger family of cheap hardware where connectivity is bolted on rather than included, which lowers the cost of taking part.

404-GEN puts its 3D model generator inside Blender

404-GEN, the subnet that turns a written description into a three dimensional model, has released an add-on for Blender, the free 3D design program used widely by animators, game makers and product designers. Instead of visiting a website, a creator can now type what they want and have the generated object appear directly in the scene they are already working on, then edit it by hand from there. The team has published an installation guide and the download alongside the announcement. What it could mean: a subnet is only useful if people reach it where they already work, and slotting into an established tool is a far shorter route to real users than asking them to learn a new one. What we would watch is whether the models arrive good enough to keep, or whether they mostly serve as rough starting shapes.

MANTIS pays out for the first time, three days after asking for Bitcoin signals

MANTIS, a subnet that collects predictive signals about the Bitcoin price, says its reward mechanism has now worked from end to end for the first time: it asked for market intelligence, saw useful submissions arrive on the opening day, and paid the winning miner within three days, with the whole process recorded on the Bittensor blockchain rather than settled privately by the team. Miners are the independent operators who supply the work a subnet asks for, and getting paid correctly is the part that most often breaks first. The team framed this as a first milestone, and the claim is deliberately modest: the machinery ran without crashing. What we would watch is whether the signals themselves turn out to have real predictive value, which is a much harder thing to demonstrate than a payout landing on time.

NIOME takes its genetic data work in front of three parts of the Scottish Government

NIOME, subnet 55, produces synthetic genomic data, meaning artificial gene sequence data that behaves statistically like real patient data without belonging to any real patient, which lets researchers work on medical problems without exposing anyone's records. The team says a relationship started months ago with Scottish Enterprise, Scotland's national economic development agency, has turned into direct access: this week it presented to three arms of the Scottish Government, among them the Department for Business, Innovation, Science and Trade. Presenting is not a contract and not a paying customer, and no money or signed agreement has been announced. What it does show is a subnet being given time in a room of national policymakers rather than a crypto audience, and the thing to watch is whether any of those meetings turns into a named pilot or a funded trial.

TensorUSD warns that any Ethereum-style version of its dollar token is fake

The team behind TUSDT, the dollar-pegged token on subnet 113, has posted a check-before-you-send warning aimed at the scam accounts that follow any new token. Two details make fakes easy to spot: the token uses nine decimal places rather than the eighteen that tokens on Ethereum normally use, and there is no Ethereum version of it at all, so anyone offering one is not them. The rest is the standard advice, worth repeating because it is what actually loses people money: confirm the contract address, the token's unique on-chain identifier, against more than one official source such as the documentation, the official account and the command line tool, and never share a coldkey, the master key that controls your funds. No theft has been reported, this is the team getting ahead of one.

TensorUSD's dollar token can now be checked coin by coin on a public explorer

TensorUSD, the subnet that issues TUSDT, a token meant to hold a steady value of one US dollar and backed by TAO, says its activity is now visible on the ViewPallet explorer. An explorer is a public website that lists everything a blockchain has recorded, so anyone can look up each time a TUSDT is created or moved, along with the contract history, the reference codes for individual transactions and the block numbers they landed in. Nothing about how the token works has changed. What changed is that a reader no longer has to take the team's word for how many exist or where they went, they can check it themselves. For a token whose whole promise is that it is fully backed, being auditable by outsiders is the part that matters, and the open question is whether the on-chain record also shows the TAO held as backing, not just the tokens issued.

Nodexo explains a model where locking its token buys you GPU time

In a long interview, Hansel Melo of Nodexo set out how the subnet's token is meant to be used rather than just traded: holders lock their tokens up, and in return they earn credits that can be spent on real graphics processing units, the specialist computer chips that AI models are trained and run on. The conversation also covered where that GPU supply comes from, the revenue it brings in and the buybacks, meaning the team using income to purchase its own token back off the market. This is an explanation of an existing model rather than a new launch, so it is worth treating as the team's own account of how things work. What it is useful for is the question it raises for any subnet token: is there a reason to hold it other than hoping someone pays more later, and here the answer offered is compute you can actually use.

The tighter rule for taking over a subnet is now merged into the code

A subnet is one of the specialist marketplaces that make up Bittensor, and each one has an owner who collects a share of what that market pays out. Until now, ownership could change hands when the combined stake of several separate accounts, measured as conviction, meaning stake deliberately locked in place, passed 10 percent of the subnet's eligible tokens. That let unrelated holders be counted together, including stake belonging to the current owner, so a takeover could be triggered by people who were not acting together at all. The rewritten rule, previously only a proposal and now merged into the code, requires a single account to hold conviction above 18 percent on its own before it can take a subnet over. It sits inside network upgrade 447, which is built but still waiting on the second of the two signatures needed to deploy it, so what changed today is that the code is settled, not that the rule is live yet.

The old command line tool now warns on its download page that it has been replaced

Bittensor's command line tool, the program people type commands into to manage a wallet and move stake, was reorganised a while back: the separate bittensor-cli package was folded into the main bittensor one. The problem was that the download page for the old package, on PyPI which is the public library that Python programs are fetched from, gave no sign of that. A warning had been added, but it was written in a format only GitHub understands, so it showed up as raw punctuation instead of a notice. Release 9.23.2 rewrites it in plain text with links to the current package and the documentation. It changes nothing about how the network itself runs, but it does mean someone searching for the tool is less likely to install the wrong one and then wonder why it is out of date.

Chutes puts a brand new open AI model up for rent the same weekend it appeared

Chutes, the subnet that lets anyone rent computers to run AI models, is now serving Qwen3.8-27B, a model whose inner workings were released publicly by the Qwen team on Friday under a licence that lets anyone use it commercially. Chutes says it can accept text and images, hold roughly 262,000 tokens at once (a token is a chunk of text, so that is a few hundred pages of reading in one go), and costs about 40 cents per million tokens of text fed in. The pattern here is the story rather than the model: a subnet that can host a major new open model within days of its release is the argument for renting compute from a network instead of from one company, and it is worth watching whether the speed holds the next time round.

The code search on Bittensor's developer site stops breaking

The search box that lets developers look through Bittensor's own source code had been failing outright, returning a server error every time it was used, because it tried to read a copy of the code that was not actually present on the machine serving the page. A fix has now been merged: a searchable snapshot of the code is built in advance and shipped with the site, so the search answers from a file it definitely has. This changes nothing for holders and nothing about how the network pays out, but a broken code search on the official developer site is the kind of small friction that quietly puts off the people who might build on it.

Moving stake between subnets can now be capped at a price

When you shift your stake from one subnet to another, the price you get can move between the moment you press the button and the moment the change goes through. Buying and selling already had a protection against that, a limit you set so the deal only happens at a price you accept, but simply moving stake did not. That has now changed in the code: a move can carry a minimum price, and you can choose whether it fills partly or not at all if the price slips past your limit. We reported this as a proposal, and it has now been merged, which means it is written into the software and waiting on the next network upgrade to reach the live chain.

Macrocosmos says its scattered training run has passed 100 billion tokens

Macrocosmos, the team behind several Bittensor subnets, says its Orion model has now been trained on more than 100 billion tokens, the small chunks of text a model learns from, and is still improving. The interesting part is not the size of the model but how it was trained: spread across many ordinary machines in different places rather than one tightly wired data centre, which the team claims makes it the largest model ever pretrained that way. Training normally demands thousands of identical chips sitting side by side, so proving it can be done on scattered, mismatched hardware is exactly the thing a network built out of strangers' computers needs to demonstrate. The claim comes from the team itself and there is no independent benchmark of the model yet.

The overcharge on claiming root rewards is fixed in the code

Claiming rewards from root staking, the oldest and simplest way to back the network, shows you a fee estimate before you confirm. That estimate assumed far more work than the job actually involves, because the network cannot hold more than 128 subnets and the quote was not written against that ceiling. The correction has now been merged rather than merely proposed, so the number people are shown should match what the work really costs. Nobody was losing money on the difference, but a quote that is wildly too high puts people off doing something perfectly ordinary.

Linked limit orders are now written into the code

A limit order is a standing instruction to move stake only at a price you name, rather than at whatever price happens to be there. The second version, now merged, lets one order be sized from what an earlier order actually produced, so instructions can be chained: do this first, then put a set share of whatever came out of it into the next step. Orders written the old way keep behaving exactly as they did. What has changed since we last mentioned this is that it is no longer a proposal under discussion, it is in the software.

The chain now keeps a running total of the stake in each subnet

Every subnet issues its own token, called alpha, and people lock that token up to back the operators working in that subnet. Until now the chain kept no single figure for how much alpha was locked in a given subnet, so anything that needed the number had to add up thousands of separate entries one by one. The change adding that running total has moved from proposal to merged. It sounds like bookkeeping, and it is, but it is the kind of figure wallets, dashboards and future rules all lean on, and totting it up by hand is both slow and easy to get wrong.

Bitstarter switches on its own validator

Bitstarter, which helps new teams get started on Bittensor, has turned on its own validator. A validator is one of the machines that judges the work being done on the network and earns a share of the payouts for doing it. Bitstarter says it will use that position to help shape how new teams are scored from their first day, share what it earns with those teams, pay something back to people who stake with it, and put its fees into the tokens of the teams it backs. The validator is live now. How much any of that recycling is actually worth depends on how much stake it attracts, which is worth watching rather than assuming.

Apps built on the Ethereum bridge can now read a wallet's staking list

Bittensor can be used from apps written for Ethereum, the other large blockchain, through a piece of translation software known as a bridge. Those apps had no simple way to ask which accounts a given wallet had staked with, so showing a person their own position was harder than it should be. A merged change adds that lookup, handed over in pages so a very long list does not choke the request. This is plumbing for developers rather than anything a holder will see directly, but it is the sort of gap that quietly decides whether outside wallets and dashboards bother to support the network.

Bittensor's pitch contest says it raised over 800 TAO and sent two teams towards launch

Bitstarter runs Pitchtensor, a live contest where teams pitch an idea for a subnet, one of the specialist networks inside Bittensor that pays people to do a particular kind of AI work, and the room funds the ones it believes in. The organisers say the first round, held at the Proof of Talk conference, raised more than 800 TAO and that two of the subnets pitched there are now building towards going live on the main network. Applications are open for a second round at the Exploit Summit, this time limited to machine learning teams. The figures come from the organisers themselves, and a team heading towards the main network is still a long way from a working product.

TensorUSD sets out how its TAO-backed dollar token is made and unmade

TensorUSD, which runs as subnet 113, described the full life of TUSDT, its dollar-pegged token. Someone locks TAO worth 150% of what they want to borrow, meaning 150 dollars of TAO to mint 100 dollars of TUSDT, then spends or holds that token while interest quietly builds on the debt. To get the locked TAO back they repay the original amount plus the interest, and the original amount is then permanently destroyed rather than recirculated, which is what burning means here. The design point they are making is that the token is only ever created against real collateral and shrinks again when the loan is settled. This is an explanation of how the system already works rather than a new feature, so the thing worth watching is whether the amounts actually minted and repaid grow, not the description itself.

The repair to the subnet ownership counters is now merged

Every subnet keeps a running count of its own token, known as alpha, and those counts decide how much backing someone needs before they can take a subnet over. Two of those counters had drifted away from reality over time. The repair was written up earlier this month as a proposal and has now been merged into the code, arriving as part of network upgrade 446, so this has moved from being discussed to being agreed. It only corrects the bookkeeping: it does not create new tokens, move anyone's stake, or undo past transactions. It also narrows the pool of tokens that counts towards a takeover attempt, leaving out holdings that could never support a challenger anyway, which means the bar for seizing a subnet is now measured against a smaller and more honest number.

Leftover data from before last year's big rebuild gets swept off the chain

Every machine that helps run Bittensor keeps a full copy of the network's records, so anything left lying in there costs disk space and time for everyone, forever. A change now merged into the code clears out two kinds of junk: entries from before dTAO, the 2025 rebuild that gave each subnet its own token, and rows that were only ever storing a zero or a default value. It is written to work in small bounded batches and remember where it got to, so the sweep does not stall the network while it runs, and anything holding a real non zero value is left untouched. This is housekeeping rather than a new feature, and nobody using Bittensor will see anything change, but it makes running a node slightly cheaper and slows the steady growth of the chain's stored history.

Mantis switches on new rules for paying its traders

Mantis, the subnet where people compete by submitting trading signals, says it has launched an upgraded incentive mechanism called FLOW. The incentive mechanism is simply the rulebook that decides which contributors get paid and how much, so changing it changes who earns. The team says the new rules reward anyone who can trade precisely and stay alive through market swings, and that the next step is opening proprietary trading evaluations, the tests funded trading firms use to hand out capital, so the subnet's combined signal can be put behind real money. What to watch is whether the promised capital access actually arrives, and whether rewarding survival rather than raw returns changes who ends up at the top of the board.

Chutes puts a newly released open AI model live within days

Chutes, the subnet that rents out computers to run AI models for anyone who wants them, says it is already serving Qwen3.8-27B, a model whose innards were published by the Qwen team on Friday under a licence that lets anyone use it commercially. It takes text and images, can hold a very large amount of material in view at once, roughly the length of a long book, and Chutes prices it at forty cents per million tokens of input, tokens being the small chunks of text an AI reads. Chutes also says it runs inside hardware that can prove what it is running, with the traffic encrypted end to end, meaning the operator should not be able to read what you send. Turning models around this quickly is normal competitive housekeeping for an inference provider rather than a change in what the subnet does.

Macrocosmos says its 16 billion parameter model has passed 100 billion training tokens

Macrocosmos, the team behind several Bittensor subnets, says its Orion training run has passed 100 billion tokens, tokens being the word fragments a model reads while learning, and that quality is still improving. The point of the run is where it happens: instead of one company's data centre, the model is being trained across ordinary graphics cards rented from strangers all over the world, machines that differ from one another and can drop out at any moment. The team calls it the largest model ever pretrained this way, and says it will add new card types and raise the machine count to 256 next week. That claim is theirs and has not been independently checked. What it could mean is that training a large model without a single owner's hardware becomes a real option rather than a demonstration, though the test is whether the finished model holds up against ones trained the conventional way.

Subnet 88 says its quant fund and public index are both running

Investing88, the team behind subnet 88, laid out its record since launching in April 2025: an 88 Quant Fund opened at the end of 2025, followed by an 88 Public Index at the end of March, with the index accepting a maximum of 10,000 TAO of customer money. Tao.com repeated the claim, calling it evidence of progress beyond roadmap promises. Most of this has been said before, so the news here is the detail rather than the direction: a stated cap on how much the index will take, and the claim to be among the small number of subnets earning money from customers rather than only from network payouts. None of the figures have been independently verified, and the team has not published what the fund or index has returned.

Nearly half a million doomed transactions a month now get turned away earlier

Every action on Bittensor is written into a block, a permanent page in the network's shared record, and until now some actions that were certain to fail were still being written down before failing. A merged change counts the scale of it: in the last thirty days, 448,600 attempts failed simply because a validator, one of the participants that scores the work miners do, submitted its scores sooner than the rules allow, plus a second batch that failed because the account submitting scores was not permitted to submit them for that subnet. Both are now checked before the attempt is allowed into a block, so they are refused up front instead of consuming space in the record. Nobody's pay changes and no new feature appears: the effect is a slimmer chain and faster, clearer rejections for anyone running validating software.

TensorUSD spells out what winning one of its auctions pays

TensorUSD, subnet 113, is building a dollar-pegged token called TUSDT that people borrow against assets they lock up, and when the locked assets fall too far in value the position is auctioned off, a process known as liquidation. The team set out exactly what a winning bidder receives: the locked collateral minus a 0.3% fee to the network, while the money bid repays what was borrowed, the borrowed amount itself is destroyed rather than recirculated, the interest goes to the project's treasury, and the bidder also earns TAO rewards paid out by the subnet for taking part. It is a description of how the system works rather than a new launch, but it is a specific one, and it shows the design intends every auction to shrink the supply of the dollar token and feed its treasury at the same time. What we would watch is whether real auctions actually attract bidders, since a lending system with no one willing to buy the collateral is where these designs usually come unstuck.

Ditto rewrites its scoring test after miners learned to game the old one

Ditto, subnet 118, pays the people running its machines according to how well their AI agents score on an in-house test called Dittobench. The team says version 10 was maximised within hours, meaning competitors found ways to score highly without their agents actually becoming more capable, a problem known as overfitting, where a model learns the answers to the test rather than the underlying skill. Version 11 is now live and is meant to close that gap, and the team says coding ability is the next thing it will measure. Rewriting the test is routine for a subnet that pays for measured performance, but each rewrite reshuffles who is winning, so the question is whether scores keep climbing once the test stops moving.

ByteLeap says subnet 128 is being paid by the network again

ByteLeap, the team behind subnet 128, says the network has switched its emissions back on. Emissions are the stream of newly created TAO the Bittensor network hands to each subnet every day to pay the people running and checking its work, so a subnet that loses them is left funding itself. ByteLeap also acknowledged that it had stopped keeping its community informed, and said it will begin publishing revenue and other figures from the business behind the subnet so outsiders can see how it is doing. The restored payout is the concrete part; the transparency is so far only a promise, so what we would watch is whether a real revenue number actually appears, and whether it keeps appearing on a regular schedule rather than once.

Score says other Bittensor projects are now plugged into its video studio

Score, the project that uses AI to watch sports footage and label what happens in it, says two other Bittensor projects are now wired into Score Studio, its editing and tagging tool: Hippius, which sells decentralised storage and cloud capacity, and Say GM. Alongside that it says it has a working commercial arrangement with Vidaio, another video project on the network. This is worth noting because subnets, the independent teams that make up Bittensor, mostly sell to the outside world, and one of them paying to use another one's service is a small sign that the parts of the network are starting to buy from each other. The detail here comes from Score's own side and the other teams have not set out their version of it, so the arrangement is their account rather than a jointly announced deal.

Leftover data from removed subnet slots is now cleaned up

Every participant on a subnet holds a numbered slot, and attached to that slot they can leave small pieces of data on the chain: sealed submissions waiting to be opened, notes, and the deposits held against them. Until now, when a subnet reduced the number of slots it offers, the participants pushed out left all of that behind permanently. A merged change adds a proper clean-up step so the chain stops carrying records for people who are no longer there. Ordinary slot swaps, where one participant simply replaces another, work exactly as before. It is housekeeping rather than a new feature, it does not change anyone's pay or stake, and it takes effect with the next network upgrade.

Shared wallet transactions stop being charged for work they did not use

Bittensor lets one account act on behalf of another, a setup called a proxy, which is how shared wallets, team accounts and delegated signing work. Until now, when a transaction was sent through a proxy the network charged it for the worst case amount of work the instruction could possibly take, and then quietly threw away the refund when the real work turned out to be less. Direct transactions got that refund, proxy ones did not, so anyone using a shared wallet was paying more in fees than they needed to and taking up more room in each block. This fix passes the real cost back so proxy transactions are billed for what they actually used. It is merged and part of the code now going out. Small on its own, but it is the kind of quiet accounting error that makes the safer, shared way of holding an account cost more than the simple one.

The network stops docking pay from subnets that hold back miner rewards

Every day the network prints new TAO and splits it between subnets, and each subnet is meant to pass a share of that on to its miners, the people running the computers that do the actual work. A recent rule shrank a subnet's slice of the network's payouts if it burned that miner share, meaning destroyed it rather than paying it out, a trick some teams use to support their own token. Upgrade 444, now merged and carried forward into the later upgrades, removes that weighting entirely: a subnet's share is decided by its price average alone, regardless of how it treats miner pay. The same release bundles a batch of smaller repairs, including one that stopped proxy accounts, which let one key act on behalf of another, being overcharged transaction fees. The question worth asking is what is left to encourage subnets to keep paying the miners they depend on, now that withholding that money no longer costs them anything at the network level.

Subnet 88 says it has started buying its own token on the chain

Every subnet has its own token, called an alpha token, that people can buy and hold alongside TAO. The team behind Investing88, subnet 88, has said that on-chain purchases of its own token have now begun, and that the pace will pick up as a price average the buying is tied to catches up. Buying back your own token takes some of it out of circulation and is usually done to show the team is putting its own resources behind the project. No amount, budget or end date has been published, so from the outside there is no way to say how large this is or how long it runs. What we would watch is whether the purchases keep showing up on the chain over the coming weeks, and whether the team ever puts a number on them.

Tatsu gets its software running on boards with no wifi built in

Tatsu, subnet 84, builds TatsuOS, software for small cheap computers of the kind hobbyists and hardware makers use. Until now a supported board meant a board with wifi already inside it. The team took the Raspberry Pi Pico, a very cheap board with no radio of its own, paired it with an inexpensive add-on wifi module, and wrote a firmware profile, the low level software that tells the board how to behave, around that pairing. The result is that the platform now runs on a class of hardware it previously could not touch, which for a project that wants a lot of small devices online means cheaper ways in for the people supplying them. The trade is that a longer list of supported hardware is a longer list to keep working, so the thing to watch is whether the older boards stay well supported as new ones are added.

Hardware wallet owners can now read a trade before they approve it

A limit order is a standing instruction to buy or sell a subnet's token only at a price you pick. Until now, anyone using a Ledger, the small physical device people keep their coins on, had to approve one of these by signing a long meaningless string of letters and numbers, so the screen never told them what they were actually agreeing to. A change merged into Bittensor's core code lets the device display the order in ordinary readable words instead. Nothing is taken away, the old method still works alongside it, so this is a safety improvement rather than a new feature. It matters because approving something you cannot read is one of the easiest ways to lose money by accident.

Score shows its studio labelling a raw sports clip by itself

Score, subnet 44, posted a video of Score Studio taking a real volleyball clip and generating the labels automatically: marking what is in each frame and separating the objects into groups, which is the slow manual step that normally stands between raw footage and a dataset a model can learn from. Teams doing this by hand can spend weeks on it. The demo is honest about its limits, the person in the video has to go back and correct the split between team one and team two, so this is assisted labelling rather than a finished dataset with no human in it. Score is positioning this against Roboflow, a widely used commercial tool with about a million users, which tells you the market they are aiming at. What it could mean is a cheaper route to custom vision datasets for anyone whose footage does not match an off the shelf model, though a single curated demo is not evidence of how it behaves on messier material.

TrendBTC says it handled 200,000 predictions in a single day

TrendBTC, an app where people guess whether the price of Bitcoin, Ethereum or Solana will be higher or lower five minutes from now, says it took more than 200,000 of those guesses in one day, the busiest day it has ever recorded. The number comes from the project itself and nobody outside it has checked it. It is also worth knowing that the app pays people in its own token, called TREND, for taking part, so a high count of guesses shows the product is live and being used but does not tell you how many people would still use it with no reward attached. The claim is a usage milestone, not new revenue or a partnership.

Broken sealed submissions are thrown out on arrival

On some subnets, miners submit their work sealed, meaning it is locked with a time based key so nobody can read or copy it until the network opens everything at an agreed moment. Some of those sealed packages turn up malformed, with the header fields the wrong size, and until now the network stored them anyway and only discovered the problem when it tried to open them. A merged change checks that both header fields are exactly the right size before anything is accepted, and rejects the rest immediately. It is housekeeping rather than a rule change: nobody's rewards move, but junk submissions no longer sit in the network's storage or get in the way of the round where real ones are opened. It lands alongside the earlier repair that stopped valid sealed work being wiped when the network failed to open it.

Root Reborn's record move is spread out so the chain keeps running

Root Reborn is the change that stops rewards from staking to Root, the network's default pool, being paid out and reinvests them instead. Switching it on means moving every existing claim record into a new format, and that move was written to happen all at once, inside the single block where the upgrade lands. A block is one step of the chain, taken roughly every twelve seconds, and doing heavy work inside one of them can slow or stall the network for everyone. This change spreads the move across many blocks and lets it resume where it left off if it is interrupted. Nothing about who is owed what changes, only how the records get there.

The step that makes a transaction permanent gets a bookkeeping fix

A change to Bittensor itself has been merged that tidies up how the network records who is allowed to declare a block final, meaning settled for good and impossible to undo. A small group of appointed signers does that job, and the chain keeps a counter saying which version of that group is currently in charge. If the counter and the group drift apart in the saved history, a computer joining the network from scratch can refuse to trust the records it is downloading. The fix keeps the two in step whenever an administrator changes the signer group, and applies a one off correction for an old mistake in Bittensor's own history at the moment a new machine begins catching up rather than later on. A companion change for the practice network developers test on has been proposed but not yet merged. Nothing about how TAO is earned or paid changes here: the effect is that new machines can join the network and check its history for themselves without needing anyone's help, which is the thing that keeps the chain verifiable by outsiders rather than only by insiders.

Sealed submissions stop vanishing when the network cannot open them

Miners on Affine, subnet 120, pay a registration fee and then send in their work sealed with a time lock, a method that keeps a submission secret until a set moment and only then opens it. Until now, if the opening step failed, the chain threw the sealed record away entirely, so a miner who had paid and submitted correctly ended up with nothing recorded and the validators who score that work saw an empty slot. A merged change keeps the sealed record in place when opening fails and writes down the failure so it can be inspected afterwards, which turns a silent loss into something someone can chase. The code is merged but rides onto the live network with the next upgrade rather than being in force this minute, so the question worth watching is whether the affected miners are made whole for the entries they already paid for.

Network upgrade 445 ships as a clear-out rather than a rule change

Version 445 of the software that runs Bittensor is out as a finished release, not one sitting and waiting on approvals like the two behind it. It is maintenance rather than new rules: transactions that are guaranteed to fail are now turned away before they are packed into a block, so they stop taking up space and burning fees for nothing; empty and out of date entries are deleted from the chain's storage; and old notes left behind by machines that have long since left the network are purged. None of this changes who gets paid what. It matters because every record the chain keeps forever has to be stored by every machine running it, so a tidier chain is cheaper to run, and cheap to run is what keeps the cost of ordinary actions from drifting upwards.

Paying straight into a Root Reborn basket is switched off while its cost is checked

Root Reborn, the change that reinvests staking rewards instead of paying them out, holds each participant's share in what the code calls a basket. Developers have temporarily closed the route that let someone deposit directly into one of those baskets. The reason is storage: each direct deposit spreads across a validator's full list of scores and can create a fresh record on the chain for every slot it touches, so they want to measure how quickly that would grow the data every node has to keep before letting it carry on. The gate covers that one deposit route, not staking in general, and it is described as temporary pending the review. What we would watch is whether the route comes back with a limit attached or stays shut.

Subnets that hold back miner pay lose part of their share of the network's payouts

Bittensor prints new TAO every day and splits it between its subnets, the separate projects that run on the network, based largely on the market price of each subnet's own token. Subnet owners are allowed to withhold some of the pay meant for their miners, the people running the machines that do the actual work, either recycling that value or destroying it. A change merged on 12 August restores an older rule that takes the withheld portion into account: a subnet's slice of the network-wide payout is scaled down by whatever it did not really hand to its miners. In plain terms, a subnet can no longer keep money back from the people doing the work and still claim the same share of what the whole network pays out. The code is merged but rides out with the next signed network upgrade rather than taking effect straight away, so the thing to watch is whether subnets that have been withholding a large share of miner pay change that once it is live.

404gen launches Atlas, a 3D production tool aimed at game studios

404gen, one of the specialist marketplaces that make up Bittensor, has launched Atlas, a platform that turns written instructions into 3D assets for games and film. Those marketplaces are called subnets: on each one, independent operators known as miners compete to do a single job well and are paid in the network's token for doing it better than the rest. The Opentensor Foundation says 404gen already works with several large game studios including Square Enix, and that a game called Colony, from the studio Parralel, goes live this month generating its assets in real time through Atlas. What makes this worth noting is the presence of a named commercial user rather than a demo, though the studio relationships are described by the subnet's own side and not confirmed by the studios themselves. The thing to watch is whether Colony actually ships this month as stated.

Numinous puts its miners head to head with the big AI models in public

Subnet 6, Numinous, has opened the Numinous Arena, a public scoreboard that runs the leading commercial AI models against the models built by its own miners, the independent operators who compete on that subnet. Both sides work under the same conditions: the same news archive to read from and the same memory of what they have already been shown, so neither group is handed better material than the other. This matters because subnets usually ask outsiders to take their quality claims on trust. A scoreboard anyone can open is a claim that can be checked, and it cuts both ways, because a bad result is just as visible as a good one.

Bitcast opens its system up for outsiders to build on

Bitcast, subnet 93, has released version 2.0.0 of Bitcast-x, the software that connects to its network. Bitcast pays online creators for posts that earn genuine checked attention on YouTube and X rather than for follower counts, with brands paying for the campaigns. The new release is pointed at outside developers: it lets someone build their own product on top of Bitcast's plumbing instead of only using Bitcast's own tools. The team is asking anyone with an idea to check it against their roadmap first, and says it expects the first outside products before long. Nothing has been built on it yet, so what to watch is whether a real product from someone outside the team actually appears.

The robotics subnet confirms the Axis Robotics tie up itself

We reported earlier that Axis Robotics said it was bringing its robot data into Bittensor through subnet 80, OpenRoboto. What has changed is that OpenRoboto has now carried the same announcement from its own account, so this is both sides saying it rather than one. OpenRoboto runs an open contest in which miners, the independent operators competing on a subnet, take a shared open robotics model and try to improve it. Real robot data is the raw material that contest runs on, which is why where it comes from matters. No figures have been published for how much data is involved or on what terms, so the size of this is still unknown.

404-GEN launches Atlas, a 3D tool aimed at game studios

The subnet known as 404-GEN, which pays a network of independent computer owners to generate three dimensional models, has launched Atlas, a workflow tool for the people who build props, characters and scenery in video games. The announcement, made by the Opentensor Foundation that runs Bittensor, says 404-GEN already works with several large studios including Square Enix, the Japanese publisher behind Final Fantasy, and that a game called Colony, from the studio Parralel, goes live this month using Atlas to generate artwork while the game is running. The studio relationships are the subnet's own account of them and none of the studios have said so publicly, so treat those names with more caution than the launch itself. What this could mean: most subnets sell mainly to other crypto projects, so a tool that ordinary games companies pay for would be money arriving from outside the ecosystem. The questions worth asking are whether Colony actually ships this month, and whether any named studio confirms the work in its own words.

Axis Robotics says it is bringing its robot data onto subnet 80

Axis Robotics, a robotics company, announced a partnership with OpenRoboto, which runs as subnet 80 on Bittensor, and said it is expanding its data work into the Bittensor ecosystem. OpenRoboto is an open contest: independent operators, called miners, take a shared robotics model and try to improve it, and whoever produces the best version gets paid in the subnet's token. The interesting part is where the training material comes from, because a subnet learning from data a real robotics firm has collected is a different proposition from one learning from whatever is already public. Neither side has said how much data is involved, when it arrives, or who can use it afterwards, so what we would watch is whether anything measurable shows up on the leaderboard or in a public model release rather than only in the announcement.

Malformed sealed submissions now get thrown out at the door

Some subnets ask miners to publish their work sealed, meaning encrypted so nobody can copy it, and unsealed later at a fixed time. A merged change makes the network check that the two header fields on a sealed message are exactly the right size before it tries to unseal anything, and reject the message outright if they are not. Badly formed submissions no longer sit in storage or get in the way when other miners' work is opened. It is a small piece of housekeeping in the same part of the code as last week's fix for lost submissions, and it mainly saves the network wasted effort rather than changing anything a holder would notice.

Network upgrade 445 is live, mostly tidying up

Runtime 445, the software every machine running Bittensor uses, has shipped. Nothing in it changes what anyone earns. It rejects transactions that were always going to fail before they take up room in a block, deletes storage entries the chain no longer uses, clears out old miner commitments, and makes proxy accounts, which let one wallet act on behalf of another, behave consistently across more actions. For an ordinary holder the effect is a leaner and cheaper chain to run, and it matters mainly as the base that the larger upgrades 446 and 447 are built on top of.

Sealed submissions that fail to open no longer destroy a miner's work

Bittensor lets miners publish work in sealed form, encrypted so it can only be opened at a set time, and subnet 120, Affine, relies on that. A bug meant that if the opening step failed for any reason the chain threw the sealed submission away instead of keeping it, so miners who had paid to register and published perfectly valid work appeared to validators as having submitted nothing at all. The fix, now merged, keeps the sealed entry when opening fails and records the failure so it can be checked afterwards rather than vanishing quietly. A second change rejects badly formed sealed submissions at the door instead of letting them clog the process. Paid-for work stops disappearing, though nobody is refunded for what was already lost.

Paying straight into a Root Reborn basket is switched off for now

Root Reborn changed Root staking so rewards are reinvested into baskets instead of being paid out. One path inside it, depositing directly into a basket, has been temporarily disabled by the developers while they measure how much data each deposit leaves behind on the chain: a single direct deposit can spread across a validator's whole list of picks and open a new record for each one, and they want to know how quickly that accumulates before letting it grow further. The rest of Root Reborn is unaffected. This is caution rather than a fault, and the thing to watch is whether the path returns with a limit attached to it.

A fix for the bug that could leave a new node stuck

Any machine joining Bittensor has to verify the chain's history before it can follow along, and it does that using a set of signatures whose version number must line up exactly. A mismatch in how that number was advanced could leave a fast-syncing node unable to prove the history and unable to catch up. The fix, now merged, keeps the signature set and its version number changing together in one step, and applies a one-off correction for the main network's early history. A matching correction for the test network is proposed and not yet merged. Nobody holding TAO has to do anything differently, but a chain that new operators can join without hand-holding is a chain more people can afford to run.

Subnets that hold back miner pay now see it reflected in what the network pays them

Every subnet receives a share of the new TAO the network issues each day, worked out from that subnet's price and demand relative to the others. A change merged on 12 August restores an older adjustment to that calculation. If a subnet's owner chooses not to pay out part of the money set aside for its miners, the people running the machines that do the actual work, that withheld portion now scales the subnet's slice of network-wide rewards down too. It applies whether the held-back amount is recycled back into the pot or destroyed outright. It closes a gap where a subnet could keep its full standing in the network while distributing less than the network assumed it was distributing. For anyone holding a subnet's token, it means the reward a subnet draws from the network lines up more closely with what it genuinely hands out.

Actual Computer offers Qwen 3.8 on the day it was released

Actual Computer, which runs as subnet 95 on Bittensor and lets people run AI models on their own machine rather than sending their work to a company's servers, says the newly released Qwen 3.8 model was available through it on release day. Qwen is a family of free-to-download AI models made by the Chinese technology company Alibaba, and being ready on day zero means users did not have to wait for someone to package it. It is a routine update rather than a change to how the network works, but it is a small sign that the subnet is keeping pace with model releases, which is most of what a tool like this is judged on.

Subnets get a choice about which scores their reward maths compares against

This shipped quietly inside network upgrade 446 and is live now. Bittensor pays validators partly on how closely their scoring of miners matches everyone else's, and a bond is the record of a validator's backing for a miner that builds up over time. The change lets each subnet pick which round of scores the bond maths compares against: the current round, the previous one, or an automatic setting that uses the previous round when the penalty for disagreeing is at its harshest and the current one the rest of the time. Automatic is the default, so subnets that do nothing get the middle path. It is a dial for subnet owners rather than something a holder can act on, and it does not change how much TAO the network hands out. What it could mean: fewer validators penalised for scores that were right when they made them, though it also gives each subnet one more setting to get wrong.

Bitcast more than doubles its buyback target for August

Bitcast, which pays social media creators to post about products for paying customers, has put a number on this month. It says 10,000 dollars of its own subnet token has already been bought back and that it intends to buy back 75,000 dollars in total during August, against the 33,000 dollars it reported for last month. A buyback means the team spends money it has earned purchasing its own token on the open market rather than issuing more of it. The 10,000 dollars is done and the 75,000 is a stated intention, so the figure worth checking is what actually lands by the end of the month. What it could mean: outside revenue, not newly minted tokens, doing the buying is rare on Bittensor, and a stated target gives readers something concrete to hold the team to.

Root staking rewards get reinvested instead of paid out

Bittensor lets you stake your TAO to the network as a whole rather than backing one particular subnet, an option known as root. A change called Root Reborn, now merged into the code that runs the network, alters what happens to the rewards a root position earns: rather than arriving as a payout, root dividends are automatically reinvested into a basket of subnet holdings on the staker's behalf. It has been prepared for the live network with a gated launch and a migration that moves the old root claim records into the new basket system gradually, across many blocks, instead of all at once. For anyone holding TAO at root this could mean ending up with exposure to individual subnets they did not personally choose, so the things worth watching are how those baskets get composed and how simple it is to step back out.

Network upgrade 446 fixes the counters behind subnet ownership

Every subnet issues its own token, called alpha, and the network keeps running totals of how much of it exists. Release 446, now merged, repairs two historical errors in those totals, which had drifted from reality, and backfills alpha that was burned before the network started tracking burns properly. It also gives each subnet a choice of how a reward smoothing mechanism called liquid alpha reads consensus scores, makes failed decryptions of time locked messages visible in the record rather than silent, and fixes a bug that stopped some nodes catching up with the chain. None of it changes anyone's balance: the release corrects the accounting rather than moving money. It matters because the same counters decide who is eligible to take ownership of a subnet, so wrong numbers there had real consequences.

Bitcast says it bought back 33,000 dollars of its token last month

Bitcast, which pays creators to post about products and sells that reach to advertisers, reports that it spent about 33,000 dollars last month buying its own token back off the open market, has already spent about 10,000 dollars doing the same this month, and intends to reach roughly 75,000 dollars before the month is out. A buyback means the project takes money it has earned and uses it to purchase its own token, which is one of the few things a subnet can point to as evidence that real customers are paying it. The team links the rising figure directly to growing revenue. The number to keep an eye on is whether the buybacks keep tracking revenue rather than running ahead of it.

dFusion launches Personas, with 750,000 people to be identity checked

dFusion has gone live with Personas, which are AI stand ins built from a contributor's own online activity: the more someone contributes, the more their Persona resembles them. Because each Persona carries reputation and earns rewards, the project needs confidence that one person cannot quietly run hundreds of them, so identity checking is handled by another subnet, Yanez, using what it calls proof of humanhood. The team says more than 750,000 users will pass through that check. It is a rare case of one subnet buying a service from another rather than building it itself, and the thing worth watching is how many of those users complete verification rather than how many are eligible for it.

First lending market on Bittensor goes live, starting with Score

MentatLend has published its design document and opened what it describes as the first lending market on Bittensor, beginning with the subnet Score. The idea is straightforward once the words are unpicked: someone holding TAO can lend it out and earn more than they would by simply staking it to the network, while someone holding a subnet's own token can borrow TAO against that holding without selling it, so their position keeps earning rewards while the borrowed TAO gives them cash to use. Borrowing against a volatile holding carries the obvious risk that a fall in price forces the position to be closed, so the terms of that, and who absorbs the loss, are the details to read before touching it. This account of the launch comes from a community member rather than the project's own feed.

IOTA passes 100 billion words of training on a shared model

IOTA is training a large language model called Orion-16B using ordinary graphics cards volunteered by people all over the world, rather than one company's data centre. It has now fed the model more than 100 billion tokens, the chunks of text a model learns from, and the quality is reportedly still improving rather than levelling off. The team believes it is the largest model ever pretrained this way. That matters beyond Bittensor: the assumption across the industry is that serious models can only be trained on tightly coupled hardware in one building, and a run this size on scattered consumer machines is evidence against it. The open question is how the finished model compares with one trained conventionally at the same scale.

Chutes adds a tool that reads documents into clean data

Chutes, which rents out AI models over the internet, has launched docuextract. You hand it a PDF or a photograph, an invoice, a contract or a bank statement, and it hands back the contents as tidy text or as structured data laid out to a shape you specify, ready for other software to use. The company says it runs inside a hardware attested secure enclave, meaning a sealed area of the chip that can prove to you it has not been tampered with, so documents are not exposed while being read. Document handling is unglamorous but it is something businesses already pay for, so it is a more direct route to revenue than most subnet features.

OpenRoboto opens a shared robotics data pool with a named partner

OpenRoboto has launched its Open Data Pool, with Axis Robotics as the first partner. The team's argument is that robotics is not held back by secretive software any more, since capable models are freely published, but by data: recordings of real robots doing real tasks are scarce, expensive and locked inside individual companies. A shared pool is an attempt to fix that by giving contributors a reason to hand their recordings over. Having one named company signed up at launch is worth more than the idea alone, and the measure of whether it works will be how many others join and how much usable data actually arrives.

NIOME says three healthcare companies are now paying customers

NIOME, which puts medical research problems in front of the network to solve, says it has taken on three commercial partners in under four months, each from a different corner of healthcare. The first named one is Floré, a company that builds personalised products around a person's gut bacteria. The point the team is making is the one that matters most about any subnet: whether anybody outside the ecosystem is actually using it. Three real customers is a modest number in absolute terms but a meaningful one for a subnet this young, and the question worth asking is how much each of those relationships is worth and whether they renew.

Investing88 opens a capped public fund and routes a fifth into buybacks

Investing88, which runs investment strategies on the network, has opened what it calls the 88 Public Index with a deposit ceiling of ten thousand TAO. A fifth of what comes in is automatically used to buy the subnet's own token on the open market. The team says it is one of the few subnets already earning real money, expects deposits to grow substantially over the next year, and describes its work on Bittensor as a testing ground while traditional finance, particularly US shares, remains the main target. A cap on deposits is unusual and generally means a strategy that does not scale indefinitely, so the sensible question is what happens to returns as the fund fills up.

Nodexo passes 100 verified machines across 16 countries

Nodexo, which assembles rentable computing power from independently owned machines, says it has crossed one hundred verified graphics cards, spread over seven different card types and sixteen countries, with uptime of about 99.8 percent over the previous day. Verified here means the subnet has checked the hardware really is what its owner claims, which is the recurring problem with renting compute from strangers. It is a milestone rather than a change in what the subnet does, and the meaningful follow up is how much of that capacity is actually rented rather than sitting idle.

Fix stops sealed submissions vanishing before they can be read

Some subnets ask participants to submit work sealed, so it cannot be read until a set time has passed, which stops anyone copying a competitor's answer. A bug meant that if the network failed to open one of those sealed submissions on the first attempt, it threw the sealed copy away entirely, so the work was simply lost. On subnet 120 that meant participants paid to register, submitted valid work, and the checkers saw nothing at all. The merged fix keeps the sealed copy when opening it fails, so it can be tried again. It affects a small number of people but it was costing them money for nothing, which is the worst kind of bug.

Actual Computer offers a new open model on release day

Actual Computer, which lets people run AI models on their own machine, has made Qwen 3.8 available the same day it was published. Qwen is a family of freely available models from the Chinese company Alibaba, and the mid sized version has been well received by early users. Day one availability is a small thing but a telling one: it shows the plumbing for adding new models is quick rather than a project each time, which is what turns a tool like this into somewhere people keep a working library of models.

Engy starts buying back its token with customer money

Engy, which runs as subnet 53 and sells verified AI answers (customers pay to run large open AI models and get proof the answer really came from the model they asked for), says it has taken in more than $150,000 in user deposits and has begun using that income to buy back and destroy its own token. Alpha is the token each subnet issues, and burning means permanently removing coins from circulation so fewer exist. Engy reports 1,171.50 alpha burned across two buybacks, worth roughly $7,354 at the time of its update, and says the programme will now scale with actual usage instead of being a fixed amount funded by newly issued coins. It matters because it links the token to money arriving from customers rather than to tokens the network prints, though the sums so far are small.

Bitcast more than doubles its monthly buyback to $75,000

Bitcast, subnet 93, pays social media creators to promote products and takes a cut of what advertisers spend. It says monthly revenue keeps growing and it is scaling its buybacks to match: $33,000 of its own token bought back on the open market last month, a $75,000 target this month, with $10,000 of that already done. The team says it is profitable rather than merely revenue generating, meaning it collects more from customers than it pays out to the miners doing the work, and that 19% of the subnet's token supply is locked rather than sellable. Spending real income on your own token is unusual here, where most projects live on newly issued coins, so the thing to watch is whether these revenue figures hold up across several months.

Chutes moves its lock-up rules into Bittensor itself

Chutes, subnet 64, rents out computing power for running AI models. It is moving the mechanism that locks up its token away from a smart contract it wrote and deployed itself, and onto conviction lock, a feature built into the Bittensor network. In plain terms, the rules about who can sell and when would be enforced by the shared network rather than by code one team controls and could in principle change. Chutes says the network has grown and the machinery behind it should catch up, with further announcements promised within the week. Moving control away from a single team is generally read as safer for holders, but the detail of what the new lock actually enforces has not been published yet.

Repeated AI requests on Chutes get 90% cheaper

In the same weekly update, Chutes said the discount on cached prompt tokens has risen from 50% to 90%. When you send an AI model the same block of text over and over, for instance a long document or a fixed set of instructions at the start of every request, the provider can store the processed version and reuse it instead of redoing the work. Chutes now charges a tenth of the usual price for that reused portion. Anyone running long documents, or a chatbot that resends the same background information on every message, sees the bill drop sharply, which could make the subnet more competitive against ordinary cloud AI providers.

Leadpoet wins best product at OKX's hackathon

Leadpoet, subnet 71, says it finished first out of thousands of entries at the Genesis Hackathon run by OKX, one of the larger cryptocurrency exchanges, taking the Best Product prize. Leadpoet builds sales intelligence: software that finds and ranks the companies most likely to buy what you are selling. Through the OKX partnership it has connected that to AI agents, meaning automated software that can go hunting for customers without a person driving it. A hackathon prize is not a signed customer contract, but it is one of the rarer cases of a subnet being judged by an outside company rather than by the Bittensor community talking about itself.

A robot running and jumping contest opens on subnet 49

Nepher Robotics, subnet 49, has opened its Humanoid Run and Jump Tournament, where anyone can submit control software for a simulated two legged robot and be scored on speed, jump height and staying upright. The task was completed earlier in the week, then connected to the subnet's environment library with the scoring rules finalised. Nepher says it is deliberately using small, lightweight models combined together rather than one very large general purpose model, arguing that fast physical skills need speed more than breadth. This is a research competition rather than a product with paying customers, so the question worth asking is whether winning entries transfer to real hardware.

Subnet 46 changes direction and renames itself Instant

Zipcode, subnet 46, is pivoting into inference, the business of actually running trained AI models to answer requests, and is renaming itself Instant. The move was described by Mark Creaser, an early backer, after a call with the team. He frames it as a reaction to recent changes in how Bittensor decides which subnets get paid, saying a number of otherwise good teams are discovering their plan was written for a network that no longer exists. A pivot this sharp reads two ways: a team responding honestly to reality, or an original idea that never found customers. Only the next few months of delivery separate the two.

Score opens up the work it grades in private

Score, subnet 44, has put the second part of its private track audit onto its public console. Subnets grade the miners working for them, and when the test questions are kept secret nobody outside can tell whether the grading was fair or rigged. Score now publishes once a day the challenge that was set, the best miner's answer and the correct answer, side by side, for anyone to inspect. It is a small piece of plumbing, but being able to check a subnet's own marking is one of the harder unsolved problems across Bittensor, and this is a concrete attempt rather than a promise.

SOMA ships two tools that shrink what you send an AI

SOMA, subnet 114, published a progress list marking two pieces as shipped. SOMARIZER compresses plain text before it goes to an AI model, with a dial for how aggressive the squeeze should be. The second tool compresses the running record of an AI coding session, including the user's messages, files read, edits made and commands run, and is released as free open source software that plugs into the OpenClaw coding tool. Both aim at the same cost: AI models charge by how much text you send them, so sending less text costs less money.

Tatsu gets its sensor software onto real hardware

The Tatsu project, subnet 84, says its operating system now runs on actual ESP32 boards, the cheap chips widely used in home sensors and hobby electronics, and that setting one up no longer needs any programming knowledge. The described flow is connect, log in, install the software onto the chip, enter your WiFi details, choose which sensors you have and finish. Getting from working code to something an ordinary person can install is usually where hardware projects stall, so this is a genuine step, though the update does not say how many devices are actually out there running it.

TAO wallets can now be held on a Ledger device

The TAO Wallet browser extension from tao.com now supports Ledger, the hardware wallet maker whose devices keep the secret key that controls your coins on a small offline gadget instead of on your computer. The practical effect is that malicious software on a laptop cannot move your funds on its own, because every transaction has to be approved by physically pressing a button on the device. Support from an established hardware maker is a normal sign of infrastructure maturing. Worth noting that tao.com is also giving away three Ledger devices as a promotion, so several of the posts circulating about this are marketing rather than news.

Limit orders arrive in tao.com wallets

tao.com has switched on limit orders across its wallets. A limit order is an instruction to buy or sell only at a price you set in advance, instead of accepting whatever the price happens to be the moment you press the button. Subnet tokens are often thinly traded, meaning a single sizeable order can move the price a long way, so this gives holders a way to avoid being filled at a price they never intended. The company says more features are planned and is asking users what to build next.

Subnet 88 plans to reopen its public index with TrustedStake

Investing 88, subnet 88, says it intends to reopen the 88 Public Index, a basket that spreads a holding across many subnet tokens rather than betting on one, working with the staking provider TrustedStake. The team is waiting on Root Reborn weight setting, a recent change to how Bittensor decides which subnets receive payouts, and says it aims to be the only root validator not controlled by a single company. Nothing is live yet: this is a stated plan with a named partner, and the timing depends on a network change the team does not control.

Niome builds fake genome data so real patients stay anonymous

Niome, subnet 55, is generating synthetic genomic data on Bittensor: artificial DNA datasets that behave statistically like real ones without belonging to any identifiable person. Researchers working on precision medicine, treatment tailored to an individual's biology, need large genetic datasets, but real ones come wrapped in privacy law and consent restrictions that are slow and expensive to work around. Niome says it is built on a genomics business that has operated since 2018 and puts the market at $44 billion. This is a description of an approach rather than a launch, and the open question is who is actually paying for the data.

Chutes cuts the price of repeated AI requests by 90%

Asking the same question twice now costs a tenth of what it did. That matters for anyone running a product on top of them, because most real apps ask the same things over and over.