October 12, 2026 · Published daily
Subnet News — October 12, 2026
SN58 Attune opens Vector mining, SN49 Nepher Robotics starts a Go2 maze-navigation tournament, and SN65 True Performance Network announces its first release-ready optimized model. SN118 Ditto changes its mining direction and interim emission allocation. Also: SN53 Engy referrals, the taostats API migration, scheduled launches, incident follow-ups and a directly verified SN112 owner-key change. Refreshed through October 12 at 11:58 AM Eastern, including SN112 THOR’s new identity, Const’s SN65 commentary, additional verified code changes and Yoko’s pre-launch selection.
- Coverage window
- Oct 11, 2026, 2:17 PM – Oct 12, 2026, 11:58 AM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 22
Biggest stories first
Today's ranked news
taostats API migration: October 8 launch, November 9 retirement deadline
Late-found infrastructure update: taostats’ primary launch article is dated October 8 and says the new API is live, with the old /api/ endpoints switching off on November 9. An October 12 X reminder says 30 days, but the published migration timeline gives the precise date. Existing keys, host and Authorization header stay the same; paths move to /v1/, with revised endpoint names and sorting parameters. Developers should follow the guide and test field mappings. The provider’s speed and accuracy claims were not independently benchmarked here.
SN49 Nepher Robotics opens a Unitree Go2 LiDAR maze tournament
SN49 Nepher Robotics announced a live Unitree Go2 maze-navigation tournament on October 12. Miners train a high-level planner using only a 360-degree LiDAR scan, above a frozen walking policy, with no camera or map. The team provides an Isaac Lab environment, training/evaluation scripts and fixed benchmark mazes. It advertises a $13,900 prize pool split 70/20/10 among the top three, with submissions closing October 18. The important step is moving the navigation challenge from a wheeled vehicle to a quadruped; a simulated benchmark win is not proof of dependable deployment in every physical environment.
SN58 Attune opens Vector to miners training adaptive models
SN58 Attune’s October 12 announcement opens Vector: miners can train models and submit weights to compete for rewards. Its public documentation describes adaptive robot intelligence that learns from examples at deployment while model weights stay frozen. This puts a concrete mining competition behind the project’s in-context-learning ambition. Vector is open; the separate Horizon track remains coming soon. Participation does not establish real-world robot deployment or guaranteed rewards.
SN65 True Performance Network announces TPN005 as its first release-ready model
SN65 True Performance Network says TPN005 reduces the size of the Ministral-3-8B base model by about 40% while retaining its full 32k context. The October 12 announcement calls this the first successful upgrade and the first competition scored entirely on Fluid Bench, and links the Competition 5 dashboard. These are the team’s results, not an independent replication or proof that every capability is unchanged. The full numerical case study and consumer product are still forthcoming; release-ready does not by itself mean a public downloadable model or launched consumer service. At 15:39 UTC, Const quote-posted the announcement with “Bravo to the newest neo lab on Bittensor.” That is ecosystem commentary, not independent reproduction of the model results.
SN118 Ditto ends its memory track and announces an interim emission allocation
SN118 Ditto’s October 12 post says its memory track is finished and a router competition is next. A reranker, learned retrieval weights and a small language model under seven billion parameters are also planned. Until the new incentive mechanism is live, the team says miner emissions go to its UID 21 to fund development. That allocation is a team statement here, not a separately reconstructed chain-payment total. This is a consequential change for miners: the replacement competition is not yet live, and the team’s claim that memory is solved should not be read as a universal scientific result.
Yoko opens selection for its first subnet bouts; ForeverMoney points to TAO bridging
Yoko’s October 12 post asks users to select three subnets for its first prediction-market bouts and says the podium locks Sunday at 20:00 UTC. ForeverMoney quote-posts that invitation, explicitly saying the market has not yet gone live and pointing to existing TAO bridging to Robinhood Chain through its site or Talisman. This connects a planned Bittensor-focused application to available cross-chain infrastructure. The invitation is verified; completed trading launch, final selected subnets and volumes are not. This is reported as product planning, without a recommendation to trade or bridge funds.
SN17 404 extends the community vote and opens it to all GitHub users
SN17 404’s October 12 Game Jam rules extend community voting by 48 hours to October 14 at 10:00 UTC, or 6 AM Eastern. Voting is now open to anyone with a GitHub account, rather than entrants only. The rules allow one account, one vote on a shortlisted entry; entrants cannot vote for their own game, and podium winners cannot also take the community prize. The published community prize is 0.5 TAO. This changes who can participate and when voting closes, rather than announcing a new main competition winner.
SN22 Desearch v0.0.236 changes page checks after false-lockout concerns
SN22 Desearch’s new published validator code moves to 20 checked pages per 1,000-URL task, fails a task at five or more mismatched pages, removes JavaScript-rendered rechecks and adds an hourly ScrapingDog budget. Unconfirmed fetch errors alone do not fail a task. This supersedes yesterday’s rendered-recheck implementation, so operators should use the current version rather than combine incompatible rules. The dated digest reports earlier false lockouts and a rollback; the new code is verified, but universal deployment and resolution for every affected miner are not.
SN51 Lium fixes edits to pods rented from public templates
SN51 Lium’s October 12 SDK change fixes HTTP 400 errors when editing a pod rented from a public template, a regression described as present since version 0.15.0. The request now omits the temporary-template field while preserving one-time-template cleanup when enabled. For GPU renters, this addresses a concrete management failure after renting a pod. The merged fix is verified; a published package version and universal rollout are not established by this commit alone.
SN53 Engy launches referrals paying a share of referred API spending
SN53 Engy announced its referral program on October 11, after the preceding report’s cutoff. The official post offers 10% of referred customers’ API spending and directs users to the referral dashboard. The significance is distribution: people bringing paying inference users can participate in customer acquisition. This is a referral commission, not a subnet staking yield, guaranteed income or proof of a new enterprise contract. Detailed eligibility, duration and withdrawal terms should be checked in the program rather than inferred from the short announcement.
SN64 Chutes brief reports TEE v1.5.0; AMD hosts still restricted
SN64 Chutes’ October 12 digest reports TEE v1.5.0 with AMD SEV-SNP support, serving-preserving maintenance and a TDX reboot-hang fix. Crucially, it says SNP hosts are not yet accepted because of open security bulletins. It describes an upgrade window through October 25 and up to two servers per miner upgraded concurrently. These details come from a dated digest whose private originals were not independently read. Implementation support is not admission approval, and the discussed future model additions are not verified live launches.
SN71 Leadpoet tightens company-name verification and updates its reveal decoder
SN71 Leadpoet merged two operational changes on October 12. Company-fit scoring can now reconcile legal and trading names when first-party wording and an exact independently fetched company profile support the same identity; denied or merely proposed aliases are rejected. A separate change adds decoding for measured Finney runtime 477 weight-reveal events. Together these address false identity mismatches and compatibility in reward-related event handling. The code is verified, but improved sales conversion, live validator adoption and a network-wide runtime upgrade are not inferred from these merges.
SN78 UMI preserves signed references when replaying paid-service evidence
SN78 UMI’s October 12 merge preserves the exact committed bytes of retained service references through certification replay. Before opening new paid-service work, the lifecycle checks selected terms, signed catalogs and reference files; missing or changed inputs hold progression, including after a restart. Already-running work retains its existing recovery rules. This strengthens the evidence behind certification and settlement without rewriting signed commitments. It is a verified implementation change, not proof of new customer revenue or a completed app launch.
SN89 InfiniteQuant carries unpaid claims forward and adds finite market depth
SN89 InfiniteQuant’s October 12 code schedules unpaid reward claims to carry into later cycles from 01:00 UTC rather than being dropped above the cycle’s emission budget. A second change, armed for 05:00 UTC, gives each account $100 and each entity $400 of same-side stake at model pricing across matching windows; excess stake pays a depth premium. These are code-defined accounting and pricing changes, not guaranteed payouts or proof that all validators adopted them. The current README also says USDC withdrawals are sent without manual review, while alpha withdrawals remain manually handled.
SN96 Verathos schedules its Decide transition for October 15
SN96 Verathos says it becomes Verathos Decide on Thursday, October 15. The October 12 announcement describes open-weight decision models that return a probability for each option, signed by the responding miner, with emissions following paid demand. This is a scheduled product and incentive transition, not a completed mainnet rollout. At the 14:25 UTC snapshot, its alpha/TAO price was up 16.77% over 24 hours; that is market context, not proof that the transition has succeeded or caused the move.
SN100 Cortex sets HyperTrain’s start for 18:00 UTC; proxy results retain limits
SN100 Cortex’s October 12 post sets a start for 18:00 UTC today, or 2 PM Eastern, after this report’s cutoff. Separately, its public HyperTrain repository records completion of a 150-job, roughly 1.1-million-parameter proxy grid on an RTX 5090. The documentation explicitly says no cell demonstrated parity with data-parallel training and that provisional settings remain unvalidated for a paid run. Completing the experiment is progress; it does not establish training parity at large-model scale or that tonight’s planned start has already happened.
SN105 Beam briefs report security fixes and a fresh task-volume concern
Late-found SN105 Beam coverage distinguishes two dates. The October 11 digest attributes to the team caller-ownership checks on orchestrator routes, hidden miner identities during integrity checks, corrected sale counting and lower telemetry prices. An October 12 digest then reports community concerns about a NATS protocol change, falling verified task volume and persistent debt calculations. Those later messages are not automatically team statements, and the private originals and underlying measurements were not independently inspected. These remain developing reports: earlier fixes do not establish that the later operational concerns are resolved. Operators should confirm current SDK and migration instructions with the team.
SN107 Minos bounds genomic scoring to stop late jobs holding up rounds
Late-found code update: SN107 Minos merged its scoring-deadline fix on October 9 at 03:53 UTC. New secondary work stops seven minutes before the scoring deadline by default; once results can no longer be stored, the validator stops waiting and removes remaining scoring containers. That frees CPU for the next round instead of letting a slow genomic job delay finalization. The merge and implementation are verified; adoption by every validator and a measured production speed gain are not. This is a scoring-reliability change, separate from the planned second incentive mechanism.
SN108 ChipForge brief reports revalidation, reopened submissions and 13 bans
SN108 ChipForge’s October 12 digest attributes to the team a completed re-evaluation of 85 quest_001_npu entries using hardened validation, a new winner, reopened submissions and a return to a 10% minimum-improvement requirement. It also reports 13 coldkeys banned for exploiting evaluation. The private original messages and per-account evidence were not independently read, so these remain attributed reports rather than independently proven misconduct. This updates the unresolved validation story; it does not certify every payment or submission issue as resolved.
SN112 THOR replaces its owner key and updates its public identity
A direct Finney archive-node check confirms SN112’s owner coldkey changed at 13:19 UTC on October 12, block 9,267,377. The decoded ColdkeySwapped event is attached to extrinsic 4, and owner storage differs immediately before and after that block. The registration height remained unchanged between the daily snapshots. This establishes an executed owner-key replacement, not a new registration or proof of a negotiated sale. By the 15:54 UTC market snapshot, the slot’s identity had changed from for sale to THOR, with the description Home of Open Source Robotics and a thormachines.ai website. SubnetRadar separately reports the rename. The new identity and key swap do not establish a buyer, purchase price, completed negotiated sale or launched robotics product.
SN120 Affine separates fresh training diagnostics from held-out claims
SN120 Affine’s October 12 code and documentation add a separate dashboard series for fresh generations on a fixed 128-task training cohort. Results require authenticated aggregates and complete provenance; missing or invalid observations stay unavailable rather than becoming zero accuracy. The docs distinguish the original 7B comparison reference from the actual training-run starting checkpoint, and explicitly say this is not independent held-out confirmation. Publishing the display does not establish that the producer is deployed or that the first new numerical measurement is available.
Const comments “Gamma” in an inference-for-equity discussion
On October 11 at 19:35 UTC, Const replied “Gamma” to Algod’s public question about giving startups inference instead of only cash investment. This is relevant ecosystem commentary and is included with the original thread context. The one-word reply does not announce a new protocol feature, launch date, funding program or finalized Gamma economics. Other commenters’ descriptions are not substituted for an official specification.
SN21 AdTAO plans fresh advertiser codes for each future training slice
SN21 AdTAO updated its training documentation on October 12: starting with the next slice after v4, account, campaign and ad-group codes will change between slices. Records can still be joined within one slice, but miners should no longer be able to track the same advertiser across successive slices using a stable code. Existing v4 and earlier datasets remain unchanged, and each manifest identifies its coding scheme. This is a documented forward privacy policy; the next dataset’s publication and the effectiveness of every privacy protection are not independently confirmed.
SN33 ReadyAI previews a Skills Store planned for this month
SN33 ReadyAI’s October 11 demonstration compares the same brief and models with a subnet-produced social-carousel skill, reporting large gains on its Chrome-based scoring. The post says the Skills Store is coming this month. This shows the team’s proposed route from mined skills to useful agent output, but the demonstration is not an independently replicated general benchmark and the store is not announced live. The distinction matters: a promising showcase and a dated product plan are different from a completed marketplace launch.
SN44 Score brief distinguishes resumed beverage work from a planned track pause
SN44 Score’s October 12 digest reports the beverage track resumed after roughly ten hours without work, while the fallen-person detection track is planned to be taken down for changes. These are separate challenge states, so one recovery does not imply the entire pipeline is clear. The digest also describes removing specific bad blocks. The private originals were not independently read; the resumption and pause remain developing reports and the exact maintenance timing should be confirmed with the team.
Audit trail
Coverage notes
Searched
- All 128 current subnet profiles, 92 dated briefs and 190 associated repository paths; original posts and published code checked for selected stories.
- 202 associated public X account samples, with 192 returning visible posts; signed-in X verified as @ShizzyUnchained.
- Grok daily discovery, top-five attention question and omission check, explicitly including @const_reborn; direct Const posts and replies.
- Direct finalized Finney archive-RPC owner, registration-height and network-status comparisons, with the SN112 change traced to its decoded execution block.
- SubnetRadar Moments, TAO Daily, TAO Media, AlphaGap public dashboard, partners, market outliers, emission leaders, preceding seven editions and YouTube Videos/Live.
- Checklist rerun: 128 profiles, 92 briefs, 190 repository paths and 203 public X account samples refreshed; signed-in Grok discovery/top-five/omission checks, direct Const quote and both YouTube tabs.
Public-data gaps
- X account samples and Grok are discovery aids, not an exhaustive census of every alias or post. Private Discord originals were not independently read; digest-only stories remain developing.
- The historical multi-key chain query did not return within the bounded attempt. Finalized endpoint comparisons and the traced SN112 event do not exclude a transient registration or owner change between snapshots. No exhaustive no-event claim is made.
- AlphaGap’s public response did not expose its complete live leaderboard. Published repository code does not establish adoption by every live validator. The omission prompt requested a later cutoff than the actual run; only sources observed by the report’s actual cutoff are included.
- The SN8 Vanta video post was collected earlier with a sparse caption, but its media was not reviewed then. During this correction X returns page does not exist and the post is absent from the visible profile; Grok also cannot retrieve it. Its contents remain unverified and carried forward, not treated as proof of no news. Additional media not fully inspected is not represented as exhaustively reviewed.