October 15, 2026 · Published daily
Subnet News — October 15, 2026
SN96 Verathos Decide launches with operational model endpoints, SN92 REAX identifies its decision-inference network, and SN75 Hippius launches Projects. SN28 SayGM adds video generation, SN65 TPN opens its next competition, and SN105 Beam pilots customer discovery with SN71 Leadpoet. Also: SN38 ChronoLLM’s roadmap, active runtime 479, miner fixes, unresolved incidents and the SN48 Quantum Compute replay. Coverage through October 15 at 10:31 AM Eastern; SN80 OpenRoboto’s peaq fleet integration is also confirmed.
- Coverage window
- Oct 14, 2026, 11:02 AM – Oct 15, 2026, 10:31 AM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 32
Biggest stories first
Today's ranked news
Finney runtime 479 is active
Our finalized node returns specVersion 479 at block 9,289,273, advancing yesterday’s 478. Its merged release includes an opt-out from third-party additions to a coldkey’s staking-hotkey list, with contract checks and a dedicated refusal error. It also documents rejecting EIP-7702 authorization lists above 255 entries before signature recovery. These are account/resource safeguards. The previously reduced immunity and registration-cost decay values are unchanged; this is not another halving of those limits.
SN28 SayGM adds text-to-video and image-to-video creation
SN28 SayGM’s October 14 thread announces MiniMax H3 video creation, naming SN90 KubeTEE as its provider. Users start with a prompt, one image or start/end frames to create a clip with sound. Public gateway documentation also describes asynchronous jobs, polling and MP4 output. This expands the product beyond text and still images. Demonstrations are team examples, not a quality benchmark or a guarantee of generation time. The provider relationship is attributed to the thread; a separate counterpart announcement was not obtained.
SN50 Synth fixes look-ahead contamination in backtests
SN50 Synth’s Campaign 4 patch stops a scored prompt taking a prediction that starts after it, when context could already contain part of the realized price path. Matching now looks backward within a tolerance. Other repairs use close rather than open prices and respect read-only snapshot cutoffs. This matters because a forecast must not benefit from future information. The failure class is documented; every past score and any historical ranking recalculation were not independently reconstructed.
SN75 Hippius launches Projects
SN75 Hippius’ October 15 announcement introduces separate client, team or environment spaces with their own VMs, S3 buckets/keys, CDN zones, billing wallets and invoices. It describes project-scoped access, daily costs, budget alerts and an optional block on new launches at the limit. Existing resources sit in a Default project. This makes managing multiple customers or environments more practical. The announced no-downtime project-transfer flow was not exercised in this read-only review.
SN80 OpenRoboto and peaq confirm machine identities for 30 collectors
SN80 OpenRoboto’s October 15 thread says its first 30 OR-S1 data collectors now use peaqOS Pro, each with a decentralized identity anchored on Solana. peaq’s counterpart post confirms 30live units and says future collectors will follow. This advances yesterday’s scheduled discussion/teaser to a specifically announced fleet integration, giving buyers a checkable device identity behind captured robotics data. It does not by itself prove every recording is authentic or independently validate the hardware. Connecting the fleet to machine money markets is described as the next step, not financing already received. The October 19 discussion remains a separate upcoming event.
SN92 REAX identifies its network; competitions are planned for October 16
The current SN92 name is REAX, and reax.co and its public repository explicitly identify netuid 92. The project hosts pinned decision models, checking fidelity before adding a speed component to scoring. Decisionmodels.io is its customer-facing API. The site labels inference live and schedules competitions for October 16 at 12:00 UTC; those competitions are still upcoming. Hardware tiers and shares differ across release documents, so the digest’s 80/20 claim is not treated as a verified finalized allocation. Customer volume and GDPR compliance were not independently audited. This identifies a project on an existing slot, not a fresh registration or sale. Const’s ownership denial remains context. The alpha/TAO snapshot shows +109.42% over 24 hours without attributing the move to one announcement.
SN96 Verathos Decide launches with four operational models
SN96 Verathos announced Decide on October 15: send a situation and typed questions, receive probabilities for the options and a miner-signed receipt. Its public model/status APIs returned four operational entries: kev-4b, clef-flash, d3-flash and d3-mini. Handover code specifies block 9,287,280, below our finalized head 9,289,273. This advances yesterday’s plan to a public launch with observable endpoints. No paid request or receipt validation was performed, and the legacy scoring incident is not automatically resolved. The live catalog differs from the README’s initial three-model example; today’s count uses the current API.
SN1 Apex and Bitstarter schedule an Accelerator reveal
SN1 Apex and Bitstarter describe testing incentive mechanisms with SN1 Apex miners before teams launch their own subnet, with a fuller reveal on October 16. The harness is presented as reducing competition creation from weeks to minutes; that claim was not independently timed. This is a counterpart-confirmed partnership and upcoming reveal, not proof participants have registered subnets or secured customers.
SN7 allways adds USDC routes in code and reports fee coverage
SN7 allways code adds Avalanche and HyperEVM USDC launch spokes. Its public October 14 post lists five previously available destinations; the new code is not end-to-end confirmation of both added routes. The October 15 digest attributes $2,569.95 of fees across 349 transactions to the team, described as 159% of estimated miner emissions. These are unaudited period-specific figures, not net profit after every expense or future coverage guarantees.
SN8 Vanta updates Classic rules and Instant sizes
SN8 Vanta’s October 14 published rules raise Classic portfolio buying power to 40×, with asset-class caps including 30× forex and 10× crypto. The October 15 post expands Instant Funded to $25k,$50k and $100k sizes. These are simulated-account rules, not cash in a brokerage account. A separate post advertises a $100,111.47 largest USDC reward, but the readable ledger repeats it in a snapshot labelled September 23; it is not treated as a newly executed October 14 payout, and its transaction was not independently verified. The older listing teaser still lacks venue/date confirmation.
SN10 Pareton stops closed campaigns spawning rounds
SN10 Pareton now checks open status when queuing, creating and claiming rounds. Previously closing stopped weights but could leave rounds voiding on hardware mismatch and requeueing challengers. Queued work remains available for reopening. Another correction charges short completions at full length in timing. These improve campaign integrity, not a new inference-speed result.
SN14 Cacheon announces Liquid Inference access
SN14 Cacheon’s October 14 post says it is providing inference on Liquid Inference and sees growing token volume. No revenue amount or customer contract is disclosed. Code also preserves completed evaluation stages across replacement GPU workers, while acceptance notes retain a pending GPU-runtime gate. A digest separately reports a Qwen3.8 arena and first revenue-funded buy/burn; the transaction was not independently decoded. Those claims remain developing rather than a verified recurring program.
SN15 ORO separates exhausted credits from environment faults
SN15 ORO now treats an OpenRouter 402 insufficient-credit reply as an exhausted miner run key. Runtime 3.7.1 also makes the shared task catalog thread-safe, fixing concurrent cache eviction. This distinguishes funding exhaustion from environment failure and reduces spurious failed runs. It does not establish that every anti-cheat rejection was wrong or every race is error-free.
SN18 Zeus changes reward ranking to win rate
SN18 Zeus’ merged reward function uses win rate over the window instead of mean rank. This changes the performance signal miners compete on, distinct from posting a conference talk. A merge does not establish adoption by every validator or a new forecast-accuracy result.
SN19 Blockmachine requires Base Flashblocks on every node
SN19 Blockmachine’s merged miner update requires Flashblocks on every Base node. The associated capacity-test guide documents four weighted workloads, per-workload limits and self-testing. This is a concrete eligibility/configuration change for node operators, not a new customer or proof that every node has adopted it. The channel digest is community-sourced, so the primary repository is used for the requirement.
Const highlights SN28 SayGM’s email-free signup
Const’s October 14 post, two minutes after yesterday’s cutoff, highlights signup without an email. SN28 SayGM repeats availability, and the linked public guide describes wallet-signed CLI signup and funding. This is relevant product commentary. Const’s complete-anonymity, universal-encryption and lowest-cost claims are not independent guarantees: the gateway’s documentation distinguishes confidential models running inside a TEE from frontier routes whose external provider receives the prompt under its own terms. No wallet was signed, account funded or paid request sent.
SN28 SayGM releases gm-miner 0.5.0
SN28 SayGM’s gm-miner 0.5.0 promotes dstack 0.5.12 compatibility and a rustls update. The digest reports that deprecation of the older Phala image blocked new workers. The release is verified; it does not establish that every worker has upgraded or all infrastructure faults are repaired.
TokenGauge names SN28 SayGM its 2026 provider pick
A late-found TokenGauge comparison, dated October 13, names SN28 SayGM its 2026 Provider of the Year and lists it cheapest on 17 of 21 tracked real-time model routes. The primary index uses an 80% input/20% output blend with fees included; SN28 SayGM is not cheapest on every listed route. This verifies what the index publishes, not an independent remeasurement of every bill, proof of commercial independence or a universal privacy guarantee. It is kept separate from Const’s commentary and today’s video feature.
SN30 Endure publishes a five-round forecast check
SN30 Endure reports 62 of 75 forecasts across October 4–8 were at least as large as the subsequent worst five-day drop for 15 subnets. Thirteen underestimated the drop; many others substantially overestimated it. This distinguishes coverage from calibration:62/75 is not an 82.7% exact-price prediction rate. The linked dashboard reports partial indexing, and the underlying round history was not independently reproduced.
SN38 ChronoLLM announces older vintages and a November product target
SN38 ChronoLLM’s October 14 article says Round 16 uses vintage 2025, then Round 17 begins walking backward from 2012 toward 1999, one year per round. This raises the difficulty of sourcing contemporaneous data without later knowledge leaking into training. Its first public customer product is targeted for the first week of November, with use-case contributors invited. These are planned steps, not a product already launched. The article does not adjudicate the prior Round 15 dispute.
SN44 Score teases detector building while scoring issues remain open
SN44 Score’s October 15 short video shows detection boxes on an apparent synthetic person-and-tiger clip and promises easier detector building, with more soon. The inspected visual gives context but no released specification/date. Separately, the digest attributes a pipeline pause and another failed-reproduction blacklist to the team. Cricket ground-truth paths appear in code; this does not resolve all evaluation complaints or establish misconduct beyond the attributed report.
SN54 Yanez explains Sandbox allocations
SN54 Yanez’s digest attributes lower current miner distribution to Sandbox: only KAV rewards are distributed while UAV waits for Execution. UID14 is described as a partner/referral dual-incentive allocation, with excess Sandbox emissions redirected rather than simply burned. This is a reported clarification, not a reconstructed weight/payment vector or guaranteed future miner payout.
SN56 Gradients changes training-overhead bands
SN56 Gradients replaces flat 0.75 hour overhead and lowers the task-duration minimum to 0.25 hours. A subsequent correction sets smaller-model overhead to 0.4 hours below 1B and 0.6 hours for 1–4B, retaining 0.75 hours for larger models. The intermediate first-commit values are not today’s final code. This fits scheduling to model size; it is not a measured training-speed gain.
SN65 TPN opens TPN-006 with an 8 GiB target
SN65 TPN’s October 15 announcement opens TPN-006, asking miners to shrink Ministral-3-8B Uncensored to an 8 GiB ceiling while retaining 32k context, versus the previous 11 GiB ceiling. The announced King-of-the-Hill winner receives the full 25% emission weight assigned to this competition, scored on fluid_knowledge. This is a new optimization target, not an already successful 8 GiB model or 25% of all Bittensor emissions. Yesterday’s TPN-005 evidence remains the earlier result.
SN79 MVTRX schedules v0.6.4 for October 16
SN79 MVTRX’s digest reports v0.6.4 on testnet with mainnet planned October 16. Changes include miner credit net of a weighted maker-fee share, a second asset-class pool rising 5→25%, and fundamental prices in each round’s state. Original production activation/release artifacts were not independently verified; this remains reported testnet work and an upcoming switch.
SN89 InfiniteQuant checks registration before firing orders
SN89 InfiniteQuant’s watcher checks mainnet hotkey registration before firing. Its incident note reports 38 of 2,041 fills from seven unregistered hotkeys appearing live without grading. The guard fails open when its index is stale/missing, so is not unconditional. Another docs fix says the reveal delay has been 15 minutes since August 4; that corrects stale instructions, rather than announcing a new delay.
SN100 Cortex reports AWS credits; HyperTrain remains labelled preview
SN100 Cortex’s October 14 official post reports $10,000 in AWS credits after its earlier NVIDIA Inception acceptance. Credits are development support, not customer revenue. The current HyperTrain page explicitly labels training preview/not live and estimates simulated. Public launch code does not turn a scheduled start or recap into verified scoring/payments. Grok’s initial live claim is therefore not used; actual activation and the validator-weight incident remain open follow-ups.
SN102 Connito merges its Qwen3.6 residual-expert transition
SN102 Connito’s release and current README move to residual experts beside a frozen Qwen3.6-35B-A3B base. The community-sourced digest schedules October 15 around 10 AM Eastern/cycle 17728. Public code verifies the architecture, not that every validator switched exactly then or old submissions are already excluded everywhere. Compatibility changes and fleet-wide activation remain separate facts.
SN104 TAOstatus digest reports validator queries down
SN104 TAOstatus’ digest attributes confirmation that its validator is not querying miners to a team member, without a repair estimate. Original traces and a recovery announcement were not independently obtained. This is a developing availability incident, not proof of a specified payment loss for every miner.
SN105 Beam and SN71 Leadpoet announce a pilot
Both teams’ October 15 posts announce a pilot to identify businesses and decision-makers needing large dataset transfers across clouds, storage systems and AI infrastructure. SN71 Leadpoet helps discover demand for SN105 Beam. A pilot is not a completed enterprise sale: neither side discloses contract value, conversion totals or paid transfer volume.
SN105 Beam adds external Room members through webhooks
SN105 Beam’s October 14 announcement and documentation describe connecting an external application through an HTTPS endpoint with signed deliveries, two-way communication and durable retries, without a full Beam installation. This creates a practical integration path. Signing does not mean the receiving application never sees content, and this feature does not settle the outstanding PRISM scoring complaints.
SN105 Beam digest reports a temporary pricing response
The October 15 SN105 Beam digest attributes a temporary $15/TB transfer price to the team while miner-owned-transfer scoring manipulation is addressed. Pricing is explicitly not the final fix. Debt, repayment and destination failures remain separate open issues. No endpoint identity or quantified loss is established from allegations; today’s features and pilot do not make an incident all-clear.
SN117 Everyframe releases v0.5.6 recovery and balances
SN117 Everyframe’s v0.5.6 repairs abandoned deployment preparation and adds provider-balance reporting from workers’ actual keys. This advances yesterday’s v0.5.3 recovery work. The benefit is seeing usable provider credit and recovering interrupted preparation, not proof of new inference demand or that every external provider returns a balance.
SN118 Ditto digest reports 70% burn and zero gamma
The October 15 SN118 Ditto digest says burn was set to 70% and gamma to zero, briefly raised to 100%, then restored to 70%, with an epoch or two before reflection. This advances yesterday’s plan, but finalized weights and completed recycled amounts were not independently decoded. These are reported changing settings, not a static verified policy or proof of price support.
SN120 Affine changes contested claims and records weight recovery
SN120 Affine’s fix awards contested tasks to the earliest recorded R2 claim when frozen timestamps exist; duplicates previously could zero everyone. Same-second ties use a deterministic hash, and untimestamped historical cases keep old behavior. Separate primary notes record hourly assessment freshness recovery using bounded parallel reads. This addresses allocation and stale-weight failures, without independently proving every miner’s missed reward was restored.
SN121 Sundae Bar forwards skill bundles and verifies tied leaders
Validator v1.0.24 forwards skill-bundle metadata and per-sample reference reads previously dropped from results. SN121 Sundae Bar also uses the coordinator’s resolved leader for tied entries while checking scores/margins and falling back locally on violations. This improves multi-file evaluation and consistency. The digest’s next SBC30 opening remains scheduled; no new winner is inferred.
SN122 CookingTAO reports beta revenue approaching $1,000
SN122 CookingTAO’s October 14 post says its beta platform is nearing $1,000 in monthly revenue, a team high, and schedules alpha buybacks for v2.2. This is an attributed early commercial milestone, not audited revenue or evidence the purchases executed. Product income and a future token-purchase policy are separate facts.
Chain watch: SN92 rename without changed owner/registration endpoints
Direct comparisons between finalized blocks 9,282,237 and 9,289,273 show no changes to SubnetOwner, NetworkRegisteredAt or NetworksAdded for 128 slots. Moments records SN92 Available→REAX on October 14 at 17:28 UTC; renaming is not registration or proof of sale. Whale/conviction alerts do not establish ownership. Historical range querying again returns RPC4003, so temporary intermediate events cannot be excluded. SN109 Finsight’s buyer/price and the reported SN24 multisig rotation remain unverified.
Watch the SN48 Quantum Compute replay
Yesterday’s SN48 Quantum Compute interview is a completed public replay. Player metadata shows the October 14 broadcast from 15:31:16 to 16:30:30 UTC with 59:18 recording duration. The site now removes its upcoming banner and labels the replay. Both tabs show no newer regular public/members upload verified, so the October 12 public default and October 13 members selection remain. The latest interview is available without restoring removed episodes.
Bittensor Builders House applications open for October 27–31
A late-found October 14 HackQuest announcement and its event page open applications for a five-day Bittensor Builders House near Mumbai, October 27–31. About 20 participants will be selected, applying individually or in teams of up to three, for technical sessions and project work. The organizers describe it as Bittensor×HackQuest. This is an upcoming curated builder program, not a completed hackathon, guaranteed place or evidence participating projects have launched. No application was submitted during this review.
SN33 ReadyAI previews Skill Store 33 with a limited security test
SN33 ReadyAI’s October 14 trailer says a low-cost model equipped with one subnet skill found 87% of known flaws in its test applications, across five apps and four models. The store is still described as coming. The original caption establishes a team claim and teaser, not an available marketplace or independently replicated security benchmark. The trailer was not substantively reviewed, so no additional feature or test-method detail is inferred from its media.
Audit trail
Coverage notes
Searched
- All128 current subnet profiles,98 dated briefs,193 repository paths,206 public X account samples and235 collected original posts.
- Signed-in @ShizzyUnchained: Grok daily discovery, estimated top-five attention, correction/omission queries and direct Const Posts/Replies.
- AlphaGap public pages, SubnetRadar/Moments, TAO Daily, TAO media, team/partner sources, product documentation and selected commit patches; seven previous editions.
- Direct finalized archive-node comparison of owner, registration and network-status storage; runtime/parameters and bounded historical-query attempt.
- YouTube Videos and Live tabs plus player metadata, completed-stream status, membership access and public default.
Public-data gaps
- Public samples and Grok are not an exhaustive X census. Ten account samples returned no post links; private Discord originals were not read. Digest-only items remain attributed and developing.
- Historical state_queryStorage was rejected by the public archive with RPC4003 (unsafe externally). Endpoint comparisons cannot exclude intermediate events; full-window event replay remains incomplete. Buyer identities and sale prices are not inferred.
- Some older videos and current candidate media remain substantively unreviewed. Individual viewed-media and gap outcomes are retained in the source audit.
- AlphaGap’s full paid/live leaderboard was not inspected. No paid inference, wallet transaction or independent benchmark rerun was performed. Merged code does not establish adoption by every validator.