September 7, 2026 · Published daily
Subnet News — September 7, 2026
Const the Cat and TAO rewards lead today’s five-star stories, alongside BitMind v5 and OpenRoboto’s physical-robot reward allocation. Chutes’ new CEO, Beam’s live tunnels and Zapier preview, and fresh inference results follow. Researched through 10:22 AM ET.
- Coverage window
- Sep 6, 2026, 10:44 AM – 10:22 AM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 31
Biggest stories first
Today's ranked news
Const the Cat takes off: TaoCats brings TAO pairs and holder rewards into focus
TaoCats ($CONST), the cat-themed token on StonkFun, is trading against TAO on Solana. At 10:23 AM ET its original listing showed about $1.45 million in market cap, $884,000 in 24-hour volume and 120.18 TAO paid to holders. Its stated mechanism distributes a 3% transfer tax in TAO to eligible holders with at least $20 of CONST, after an approximately 2.5% distribution/operations deduction. A separate DEX Screener snapshot of the CONST/TAO pool at 10:17 AM ET showed a roughly 1,360% 24-hour USD price rise; provider figures differ and move rapidly. This is a community memecoin, not a Bittensor subnet or a confirmed Jacob Steeves project.
BitMind releases v5: continuous King of the Hill rewards replace round-based escrow
BitMind published v5.0.0 on September 6, resolving yesterday’s pending-release status. Each discriminator lane splits rewards 85/10/5 between its current king and two predecessors; the overall allocation remains 40% image, 40% video, 4% audio and 16% generators. A challenger must improve the SN34 score by 0.01 on the same benchmark. Discriminator payouts activate only after GAS explicitly enables them and the start time passes; installing v5 alone is not activation. Generator rewards continue during warmup.
OpenRoboto allocates 85% of track emissions to two physical-robot competitions
OpenRoboto’s September 7 documentation states that the active track allocation is now 15% simulation, 42.5% real-robot π0.5 and 42.5% real-robot LingBot-VLA 2.0. Each physical track has its own seasonal pool, with 95% of that pool for the champion and 5% for qualified entries under the existing vesting terms. Previously accrued rewards are not redistributed. A shared workstation interface is also published; the documentation itself is not notice that a new submission window is open.
StonkFun reports $1.51 million in daily revenue and more than $15 million in holder rewards
StonkFun’s September 6 report lists $1,508,194 in platform revenue, $904,503 in buybacks and 6.8 million STONK burned. A separate September 7 update says cumulative rewards to ecosystem token holders have passed $15 million. These are platform-wide, self-reported figures, not revenue generated solely by TAO pairs. STONK buybacks and burns are also separate from the TAO rewards distributed by individual TAO-paired tokens such as CONST and BUTT.
Chutes appoints Christian Hyland as CEO
Chutes announced Christian Hyland as its new CEO on September 7. The company says he has spent several months helping reshape its approach to scaling and operations, and cites his background at Coinbase, Toku and First Wave AI. The appointment is a leadership change at one of Bittensor’s major inference providers; the announcement did not disclose a financing round or new customer contract.
Beam launches production P2P tunnels and previews a Zapier integration
Beam announced that peer-to-peer tunnels are live in production, connecting private infrastructure, cloud services, AI workloads and remote participants for real-time data streaming. Its new September 7 follow-up previews a two-way Zapier integration: Beam transfer, Room, Tunnel and stream events would trigger application workflows, while application events could start Beam operations. Tunnels are announced live; Zapier support is described as coming, with no release date or separate Zapier confirmation located.
Pareton reports 3.5× faster Qwen inference after its first week of competition
Pareton reports a 3.5× end-to-end improvement over stock vLLM for Qwen3.8-27B-FP8 on one H200, measured over SWE-agent traces. Its figures include median request latency falling from 727 ms to 190 ms and throughput rising from 58.7 to 221.8 tokens per second per request. The winning patch combines the model’s MTP head with fused kernels. The project reports 25 rounds, 74 submissions and 31 miners. These are project benchmark results, not a universal speed guarantee.
Lium reports 14,499 tokens per second on a single GPU
Lium’s September 6 demonstration claims Qwen3.6 35B ran at 14,499 tokens per second on one rented GPU at 85% lower cost than OpenRouter. The public announcement attracted substantial attention, but the retrieved post does not specify enough batching, hardware and pricing conditions to reproduce the comparison. Treat the number as Lium’s demonstration claim, not a per-user serving speed or a general cost benchmark.
Blockmachine adds Mantle RPC with request-based pricing
Blockmachine announced dedicated Mantle RPC support on September 7. Its product article describes a free development tier, Pro plans starting at $25 per month and request-based billing, with account-generated endpoints for existing EVM tools. This extends the SN19 service to Mantle developers. The provider’s claimed savings of up to 60% were not independently benchmarked, and offering chain support is not evidence of a commercial partnership with Mantle.
SayGM approaches 100 billion processed tokens
SayGM reported 96.48 billion recorded network tokens processed on September 7, including more than 30 billion output tokens. The total remains below the 100 billion milestone. It is a project-reported usage measure, not a count of customers, revenue or independently audited paid requests.
Zeus changes rewards for repeatedly penalized miners
Zeus merged v2.1.4 rules that assign no weight on a challenge after at least eight consecutive penalties, including missed challenges. Miners registered within the last four weeks retain a grace period. Shape errors, missing responses, hash mismatches and collusion penalties count toward the rule. The source change is confirmed; adoption by every live validator was not measured.
ChronoLLM changes Round 10 to prompt-level scoring
ChronoLLM merged Round 10 changes that score individual prompt wins instead of reducing each miner duel to one overall win or loss. The evaluation adds direct questions and new prompt categories alongside text completion. A timeout fix starts the evaluation clock after model preparation, so setup time does not consume the scoring budget.
Feval documents 12-hour evaluations with 100,000 rows per window
Feval’s current miner documentation specifies 12-hour windows sampling 100,000 rows from six pinned public datasets, with the sample seed revealed at the window boundary. New audit changes accompany the rollout. The channel discussion raises generation-budget and model-context constraints and continues to request emission activation; the documentation does not establish that live emissions have started.
TensorUSD lending code requires full debt repayment in underwater liquidations
TensorUSD merged a lending-pool change that removes residual-debt write-offs: a liquidator must repay the borrower’s full debt even when seized alpha collateral is worth less. The platform fee is waived in that underwater case, so the liquidator can take a loss. This is the alpha-collateral lending pool, separate from the vault-auction thresholds discussed on TensorUSD’s X feed. The merge does not prove deployment or that an uneconomic liquidation will attract a bidder.
Swarm’s private-submission rollout is reported live in v5.1.5.4
Following yesterday’s code merge, the September 7 channel brief and a public walkthrough report v5.1.5.4 in use. The workflow packages and checks a model, records its SHA-256 commitment on-chain and uploads it without requiring a public repository for every attempt. Crown-winning models are published; unsuccessful private submissions are not described as public. A live upload was not tested and the repository has no latest GitHub release entry, so rollout confirmation remains developing.
Graphite discussion reports a move to financial-graph tasks and fact-based scoring
The September 7 channel brief reports exclusive use of the Financial Knowledge Graph API, with synthetic validation stopped, PDF deduplication and a shift from counting submissions to counting facts. It also describes a 1,000 alpha stake requirement. A matching current primary release was not located in the mapped public repository. These operating and eligibility changes remain developing; separate misconduct allegations in the discussion are not established here.
The @buttensor account says it did not launch the BUTT memecoin
The existing @buttensor account said on September 7 that it has no connection to the memecoin and does not know who launched it. The Solana token’s community account, @BuTTensorTao, separately identifies itself as an independent community. Readers should distinguish the accounts and the token mint: 2pouN3by7twkiZGy5aEKYUpf78ALDpKRTNu2WsQkpkqt. Neither a shared name nor a TAO trading pair establishes an official Bittensor affiliation.
Ditto’s founder reports Secrets rollout and browser automation in V2 testing
Ditto founder Omar Barazanji says Secrets has rolled out and browser use is being tested in V2, extending the assistant’s ability to act across work tools using persistent memory. This is a follow-up to yesterday’s app-connections announcement, not another new integration count. Separately, current subnet code adds a native coding daemon that is explicitly disabled by default; that change should not be read as a general V2 launch.
Verathos discusses mesh fixes and offering inference access to SayGM miners
The new channel brief reports a fix for missing signed snapshots during multi-machine mesh setup, covering TLS and validator trust. A public follow-up also describes an offer of Sleipnir inference API access to SayGM miners for GLM 5.2 and DeepSeek V4 Flash. A completed cross-subnet agreement, measured customer adoption and the claimed lowest-market pricing were not independently confirmed.
Score previews segmentation models in Score Studio
Score says users will soon be able to build segmentation models in Score Studio. It has not supplied a release date in the announcement. Separately, its weekend post points readers to public consoles showing private-track challenges, leading responses and ground truth; that is an auditability reminder rather than a new console launch.
Vanta highlights its largest public reward: $100,111.47
Vanta’s September 6 post contrasts its $9 entry evaluation with a largest single reward of $100,111.47. The amount is visible on its public rewards ledger, which links crypto payments to transaction records. This is a newly highlighted historical payout, not evidence that the payment was made today or that every participant receives a comparable result.
Allways adds UNI to its supported cross-chain assets
Allways announced that Uniswap’s UNI token is live on SN7, with competing miners routing it to other supported assets. The announcement identifies the Ethereum UNI contract. This is support for the token in Allways’ routing service, not an announced partnership with Uniswap.
Gradients distinguishes H100 pricing by SXM, NVL and PCIe hardware
Gradients merged hardware-aware H100 pricing and reporting, recording the GPU product name, interconnect class and NVLink availability. The change separates configurations that should not be priced as identical machines. This is a new infrastructure update following the reward overhaul covered yesterday.
UMI adds a Mac validator deployment while translation weights remain inactive
UMI now documents running its signed Linux validator release on Apple Silicon through a bounded Docker Linux environment. This broadens the available operator setup beyond the existing native Mac miner. The project explicitly says this is not a native Darwin validator, public calibration has not started and translation weights remain inactive.
SubnetRadar retracts a misattributed Compelle self-mining claim
SubnetRadar’s September 7 correction says a non-team member’s answer was wrongly reported as confirmation from Compelle’s owners. It retracts that claim and describes new speaker-role checks and better Discord source links. A statement posted in a subnet channel must not automatically be attributed to its team. This edition does not treat the withdrawn allegation as evidence of misconduct.
Cascade adds clearer duel results and waiting-participant visibility
Cascade’s new dashboard shows verdict breakdowns by prediction horizon and the full seated field during duel-only rounds. It distinguishes miners actively dueling the king from those waiting for a later round, retaining their submissions. The horizon breakdown is explanatory: the actual verdict still uses the pooled statistic.
MicroTensor freezes 62 systems in Round 1,239 and strengthens its reward archive
The September 7 channel brief reports 62 systems frozen in Round 1,239 and planned fixes for zero-output profiles and total-task latency measurement; deployment of those scoring fixes remains unconfirmed. Separately, merged archive code ensures rewarded systems are retained even when their leaderboard state would previously have excluded them. The submission window is closed; old block deadlines are not presented as upcoming events.
Albedo discussion reports restored access to its king model
The September 7 channel brief says an access problem was resolved and the king model is now public on Hugging Face. The brief exposes original Discord message links but not a usable direct model repository link in the retrieved page. Public model availability therefore remains a reported recovery rather than independently verified download access.
Cortex’s new CLI readiness report leaves production on hold
A September 7 Cortex CLI report records a blocked staging check because access and build prerequisites were unavailable. It explicitly keeps production on hold and says no product release or deployment change occurred. This limits what can be concluded from the channel’s planned feature launch: the report is not evidence of a production outage, and a completed new rollout was not confirmed.
Capcomp fixes rewards lingering on a previous competition run
Capcomp merged a validator fix that lets a run boundary trigger a fresh weight-setting pass instead of waiting for the ordinary timer. It also refuses to pay a run the chain has already left. The change addresses stale winners remaining in force after the eligible field changes; network-wide adoption was not measured.
Iota users report inconsistent job assignment on Mac mini M4
The September 7 discussion describes TaH clients receiving no jobs on Mac mini M4 while a MacBook Air setup works. Log collection and reward-timing questions remain unresolved in the brief. These are reported individual cases; the evidence does not establish a network-wide training outage or improper reward distribution.
ORO participant reports a dashboard-only Python validation failure
A participant reports an invalid-Python-syntax error when submitting an agent through ORO’s dashboard despite local Python 3.12.1 checks succeeding. No resolution is documented in the latest brief. The available evidence concerns one submission report, not all agents.
RedTeam participants report a scoring backlog and stale dashboard
The current channel brief reports dashboard updates stalled for more than three hours while 25 commits required review. It also discusses similarity checks, but the summarized before-and-after rules are too ambiguous to restate as authoritative rejection criteria. A completed backlog recovery was not confirmed.
Bitcast consolidates its canonical YouTube subnet code
Bitcast merged a large repository consolidation bringing bitcast-v2 into its canonical public codebase. This is a source and maintenance change for the existing YouTube subnet, not a newly announced YouTube partnership or a new product launch.
Audit trail
Coverage notes
Searched
- All 128 active subnet slots and their current SubnetRadar profiles; 72 dated September 6–7 briefs reviewed, including the ecosystem channel.
- 149 mapped repository paths checked with paginated commit retrieval through 10:22 AM ET; 144 accessible. Primary diffs, current documentation and selected release states inspected for material leads.
- 128 indexed X searches by subnet name, SN number and available account aliases; 80 public account candidates checked, 79 exposing posts. 110 unique current posts retrieved from account samples and the public feed.
- Signed-in AlphaGap feed reviewed for 25 visible subnet cards. TAO Daily, TAO Media and the SubnetRadar public feed cross-checked for stories and dates.
- Outside-project checks include StonkFun, Sunrise, Raydium, Mantle, the Const account and BUTT community. CONST original token page, market pool data, transfer-tax terms and reported holder distributions inspected.
- YouTube Videos and Live tabs checked and refreshed. New September 6 members upload verified at 10:11; upcoming RL Bryer interview verified for September 9 at 16:00 UTC (12 PM ET / 9 AM PT). Existing latest public episode and Phylax replay retained.
- Rolling 24-hour subnet α/TAO prices checked; no large move was used alone to elevate a subnet story. CONST’s separate USD market snapshot is labeled in the story. News cutoff 10:22 AM ET; CONST listing rechecked at 10:23 AM ET.
Public-data gaps
- Authenticated X search redirected to login. Grok and full signed-in timelines were unavailable; public profiles expose limited samples, and indexed searches returned no relevant current results. This is not exhaustive X coverage.
- Private Discord history was not opened. AI-written channel briefs can misattribute non-team speakers, compress dates or repeat old material. Uncorroborated rollout, support and agreement claims remain developing. AlphaGap summaries were not used as standalone proof of misconduct, revenue or customer demand.
- Five mapped repositories returned 404: deprecated/deprecated, actual-computer/actual-subnet-95, Satori119/Satori, NIOME-BITTENSOR/NIOME and Oh-My-Ditto/ditto. The guessed TaoCats X account did not resolve and is not attributed to the token.
- StonkFun rewards and platform revenue are displayed or self-reported totals, not an independent reconciliation of every transaction. CONST market and volume values differ across providers and observation times. Neither a token creator identity nor a causal link to TAO’s wider price move was established.
- BitMind v5 is published, but discriminator emission activation was not independently observed. DSperse PR625 remains open and unmerged. Swarm has no latest GitHub release entry; its live-rollout report remains developing. Albedo’s model link and Graphite’s current release were not located.
- Beam’s Zapier integration is a preview, with no separate Zapier launch confirmation located. Mantle RPC support is confirmed by Blockmachine’s product announcement, not by a commercial contract. Lium’s benchmark lacks sufficient retrieved workload detail for reproduction.
- Routine changes, old launches, generic price forecasts and unsupported misconduct allegations were excluded. No verified forward alpha-unlock calendar was available. Posts after the cutoff belong to the next edition.