Dates and verification labels belong to the original edition. A past verified announcement can now be outdated. Historical project names are retained; an SN filter searches the slot across registrations. All keywords must match the story text or subnet label. Ratings describe editorial priority, not investment quality.
SN95 Actual releases toks 0.3.2 under Apache 2.0
SN95 Actual’s tokenizer has a consequential licensing change: toks 0.3.2 is its first Apache 2.0 release, allowing use in closed products at any scale subject to the licence terms. The older 0.3.0 and 0.3.1 tags retain their Business Source licence. This is more than a new speed claim: removing the earlier commercial-use restriction can make the component easier for other AI developers to adopt. The release also adds tokenizer API primitives and correctness fixes. Its 13–151× Hugging Face speed table belongs to the earlier release-candidate benchmark on specified machines; 0.3.2 does not rerun that table, and tokenizer speed is not an equivalent speedup of an entire language model. Packages are attached to the GitHub release, not published to PyPI.
SN95 Actual has a released, inspectable change that broadens commercial adoption of its tokenizer. Benchmark provenance and package availability are stated precisely.
SN96 Verathos combines GPU verification and model computation with SN106 Nodexo
SN96 Verathos announced that it has brought SN106 Nodexo’s hardware work and its own verified model-computation work together under SN96 Verathos. The described product has one fleet of verified GPUs, one payment stack and token, plus daily inference credits through locked alpha. This is an explicit operating-model change connecting hardware verification to the AI work performed on that hardware. It does not establish that SN106 Nodexo has been deregistered, sold or transferred: the current chain roster still lists it separately. No token migration was attempted, and yesterday’s reported validator recovery does not independently clear the separate web-chat and inactive-miner issues.
SN96 Verathos announces a material combination of subnet services and payment design. Product consolidation is distinguished from unverified chain retirement or ownership changes.
SN97 Albedo expands coding trajectories with SN120 Affine data and leakage filters
SN97 Albedo’s merged dataset update expands its declared trajectory sources from four to thirteen, including seven SN120 Affine-derived sources and separately pinned Open-SWE versions. It also filters benchmark overlaps and trajectories that obtained solutions from outside their own work. The evaluation sampler selects at most one rollout per task, rather than giving tasks with many recorded attempts extra chances. This is a substantive improvement to the material used to assess coding agents: more data is useful only if duplication and solution leakage do not distort scores. The code change does not itself prove a measured model-quality gain or complete deployment to every evaluator.
SN97 Albedo expands evaluation coverage with explicit sampling and leakage controls, supported by inspectable merged code.
SN102 Connito AI brief reports a first customer design; primary confirmation pending
SN102 Connito AI’s channel brief says its first customer has submitted a design through a GitHub pull request and that miners are expected to work on it next. A real customer task would be an important step beyond an internal contest. However, the exact customer design, commercial terms and completed miner work were not matched to a primary release. The directory-linked repository’s recent pull requests did not establish that claim, and the dashboard returned only a page shell at this check. Reported dashboard problems therefore remain unresolved. This is a lead to follow, not a verified paid-customer or delivery milestone.
SN102 Connito AI has a material attributed customer lead, but missing original design and service confirmation require a developing label.
SN105 Beam brief flags holding and telemetry changes; FAQ pending
SN105 Beam’s latest channel brief describes a new holding requirement and paid telemetry API, with the team reportedly preparing an FAQ and retiring the old dashboard within 24 hours. The changes matter because they can affect how miners qualify and inspect their performance. Exact thresholds, pricing, the replacement data access and completed retirement were not established from a primary release in this scan. Members’ gaming allegations are not treated as verified misconduct, and a proposed community leaderboard is not a working replacement. The public repository did not provide a fresh explanatory release during the coverage window.
SN105 Beam’s reported miner-access changes are operationally material, while missing primary terms and unresolved rollout details keep the item developing.
SN105 Beam makes its OpenClaw data-movement skill available on ClawHub
SN105 Beam announced its OpenClaw skill late October 7 Eastern, at 03:30 UTC October 8. The linked ClawHub listing is available as version 0.1.0 and describes agent-controlled object transfers and fan-out between supported storage endpoints, including Amazon S3, Cloudflare R2, SN75 Hippius and Hugging Face, plus Beam Rooms for machine-to-machine messages, streams and file exchange. This packages the network’s data movement for AI-agent workflows instead of requiring every developer to assemble the integration. The listing explicitly says private-service tunnels are not available, so the older deeper-integration roadmap is not treated as fully delivered. Availability does not establish new paying customers, revenue or measured transfer performance, and it does not resolve the separate holding and telemetry questions reported today.
SN105 Beam has an original launch announcement and reachable product listing for a practical agent integration, with unsupported roadmap features and business results kept separate.
SN108 ChipForge reaches 53 submissions and a higher design score
SN108 ChipForge’s live NPU-design leaderboard at 15:47 UTC showed 53 submissions, 26 miners and eight winner records, up from yesterday’s 42, 23 and five. The displayed on-chain winner was UID64, with score 2,067.34; the next challenger must exceed 2,274.08 under the 10% improvement rule. This is observable progress in the design competition, not another announcement that it has launched. The page still showed zero of five validators agreeing, commit-reveal delay information and no daily-reward amount, so winner records must not be treated as completed payouts. At the 15:37 UTC market snapshot, alpha/TAO was down about 16.29% over 24 hours and up 184.12% over seven days. Those returns are not USD returns or evidence of chip sales.
SN108 ChipForge has directly observed increases in participation and design score. Reward uncertainty is retained and price does not earn the rating.
SN114 SOMA makes Jev available to miners for compression decisions
SN114 SOMA says Jev is now available to all its miners for the current competition. The decision model is intended to help compressors choose what to compress and how aggressively based on context. This advances beyond the earlier merged-support notice into an explicit availability announcement. Fresh benchmark code also weights Jev input-token cost at 0.2, rather than treating the decision step as free. That number is a scoring-cost coefficient, not evidence of an 80% reduction in actual prompt size or an independently measured saving for every workload. The announcement does not resolve the separate questions about prior winner weights and payouts.
SN114 SOMA makes an announced decision-model integration available in an active competition, with matching scoring implementation.
SN122 CookingTAO targets third-week-of-October one-click SN51 lium.io mining
SN122 CookingTAO’s October 7 roadmap says SN51 lium.io will be available for one-click mining setup during the third week of October, in its beta v2 phase. The plan would extend its simplified mining workflow into a GPU-rental subnet; it is a dated integration target, not a feature confirmed available today. The announcement quotes SN51 lium.io’s separate call for additional GPU providers, but that post does not independently confirm SN122 CookingTAO’s delivery date or a formal partnership. The beta dashboard and future buyback allocation were already covered on October 7, with buybacks explicitly not started. This item adds the specific SN51 integration target and keeps those earlier milestones separate.
SN122 CookingTAO supplies a concrete, primary-sourced integration timeline relevant to miners, while availability, delivery and bilateral partnership remain unconfirmed.
Subtensor merges owner controls for proof-of-work difficulty
The Subtensor repository has merged owner controls for proof-of-work registration difficulty and its minimum floor. Previously root-only settings become configurable by the subnet owner or an Owner proxy, subject to the standard parameter cooldown and admin freeze window; root retains override authority. Owners can lower a disabled floor and set a starting difficulty atomically. This gives subnet operators more direct control over proof-of-work entry conditions, rather than forcing every such change through governance. The commit also bumps the repository’s runtime version to 476. A merged version bump is not proof that runtime 476 has activated on Finney; this scan did not independently verify chain activation.
A merged protocol change expands practical subnet-operator control, with repository implementation distinguished from mainnet activation.
Pool-lending proposal adds a disabled-by-default price-impact control
The still-open Subtensor pool-lending proposal now includes a per-subnet governance control for minimum price impact, referenced to a net 500 TAO trade. The control starts at zero, so the proposed migration does not impose a blanket 1% market move. The updated design also delivers borrowed alpha without automatically selling it or locking sale proceeds, and places ordinary funded deregistration payouts before final loan recovery. These are proposed rules, not borrowing or shorting already available on mainnet. The design specifies 25% opening market LTV with additional stressed-backing checks and 100% nominal annual interest, fixed in collateral units and collected weekly; it is not a low-interest loan announcement. Governance approval, implementation review and activation remain outstanding.
A substantial lending and market-structure proposal has materially revised terms. Its open status, disabled default and unusually high stated interest are essential context.
SN10 Pareton adds an internal Qwen 3.8 27B FP8 evaluation campaign
SN10 Pareton’s new internal campaign targets Qwen 3.8 27B in FP8 on a single RTX PRO 6000, combining quality, reliability and performance requirements under its updated evaluation setup. This continues the move toward testing useful inference systems rather than rewarding a fast response without checking whether it works. The campaign code is merged, but a configured workload is not a completed customer benchmark or a measured production speedup. The separate Opentensor discussion remains scheduled for October 8.
SN10 Pareton supplies a concrete new model-and-hardware evaluation workload, with configuration, measured results and the upcoming talk kept distinct.
SN14 Cacheon reports its Q3 proof of concept and turns toward monetization
SN14 Cacheon says its Q3 proof of concept is complete, using parallel model competitions to work on output quality, and identifies monetization as the next priority. New code lets follower publication finish after an accepted commit while a separate process monitors the later reveal, avoiding delays in signing caused by waiting for reveal status. An accepted commit still does not establish an effective revealed distribution. The combination matters for a subnet moving from technical experiments toward a product: competition results and observable rewards need clear accounting. Monetization remains a stated direction, not newly disclosed paying customers or booked revenue.
SN14 Cacheon reports a completed internal milestone and ships a relevant reward-publication monitoring change, with the commercial phase still ahead.
SN21 AdTAO adds 28-day evaluation of its Google Ads predictions
SN21 AdTAO has published a new snapshot dated October 6 that follows the same Google Ads changes at seven, fourteen and twenty-eight days. The team says roughly 500–600 new changes enter the evaluation pool each day, with younger batches reaching the shorter checks first. Longer follow-up helps distinguish a brief performance change from a sustained outcome, which matters when miners predict advertising results. This scan did not independently reproduce the results or extract a defensible numerical accuracy claim from the attached graphic. At 14:20 UTC, alpha/TAO was up 17.45% over 24 hours and 21.23% over seven days; the move is not proof that this update caused buying.
SN21 AdTAO supplies a fresh, longer-horizon product evaluation milestone. Its market outlier triggered primary-source review but does not prove causality or determine the rating.
SN25 UR publishes its launch handoff with activation still blocked
SN25 UR’s revised mainnet launch handoff still leaves the required activation identity fields unset and marks the production handoff blocked. The document separates preparatory registration and reserve arrangements from authorization to launch, and clarifies that this handoff is not an owner trim or a coldkey swap. These details matter because a registered slot and launch paperwork can otherwise look like a completed operating transition. This scan found no primary evidence that the activation gates had been passed. Yesterday’s reported large conviction-mode change remains a separate existing-alpha event, not new buying or proof of launch.
SN25 UR provides a material, explicit production gate and ownership clarification. The report preserves the blocked state rather than presenting documentation as launch.
SN28 SayGM launches customer referrals rewarded from an alpha pool
SN28 SayGM’s new affiliate program ties rewards to paying customers, rather than signups or promotional credit. A new referred customer’s first top-up begins a 30-day qualifying window. The documented pool allocates a share of SN28 SayGM alpha emissions against qualifying revenue, capped at $1 of alpha per $1 paid and reduced proportionally when the pool cannot fund that cap. This is an attempt to reward distribution that generates demand for inference. Rounds close no sooner than 24 hours; crypto-funded rewards have a seven-day hold and card-funded rewards a 120-day hold, counted after closing. Payouts are alpha stake to a coldkey. The operator supplies revenue and price inputs, and an administrator can pause or upgrade the contract; this is not a guaranteed dollar commission or an independently verified payout history.
SN28 SayGM publishes a concrete customer-acquisition mechanism, eligibility rules and an inspectable funding design tied to paid usage. Reward caps, delays and administrative trust remain visible.
SN30 Endure announces mainnet risk intelligence; Forge comes next
SN30 Endure announced its mainnet launch at 14:54 UTC on October 6, shortly after yesterday’s research cutoff. Alpha Risk V1, identified as risk.v1.subnet_alpha, has miners compete to predict subnet drawdowns, volatility, time-weighted price and liquidity depth, with predictions scored against outcomes. The goal is a risk feed other applications can use, rather than another price chart. Miner and validator participation is now invited. This resolves the earlier missing launch notice, but does not independently verify all validators setting weights or outside miners receiving rewards. Forge is described as the next integration, not a lending product already launched; initial parameter control remains with a Guardian multisig.
SN30 Endure moves from a scheduled date to a detailed primary mainnet announcement and defined prediction task. Reward evidence and the next product remain separate.
SN38 ChronoLLM reports Round 14 results and changes Round 15’s reward balance
SN38 ChronoLLM reports that its Round 14 winner won 87.2% of a 650-duel comparison against Llama 3.2 3B, 90% against Qwen 3.5 2B and 96.3% against Llama 3.2 1B. Those are results from this subnet’s paired evaluation, not a universal model ranking or an independently repeated benchmark. The team says leakage performance had stopped improving over ten rounds, exposing a weakness in its pass-or-fail incentive. Round 15 therefore gives leakage 30% of the combined score and output quality 70%, with a bounded leakage scoring window. The new balance is verified in merged code. It aims to keep the time-aware model competition useful without rewarding fluent answers that fail its leakage objective; deployment to every validator was not independently confirmed.
SN38 ChronoLLM supplies fresh scoped model results and a substantive, implemented incentive change addressing a documented weakness.
SN44 Score’s Manako goes commercial across an initial 124 Avia stations
SN44 Score’s commercial platform, Manako, has moved beyond a pilot announcement into a substantial operating deployment. Manako’s October 6 release and TaoWeave’s October 7 SEC filing say its Vision Agents now run around the clock across a first wave of 124 unmanned Avia Xpress stations in France. Existing cameras detect payment-terminal failures, damaged pumps and equipment, and camera outages, routing problems to the people who can fix them. Manako says processing happens on site without exporting raw footage, and Avia Thévenin-Ducrot is its lead Physical AI design partner. This is a concrete route from vision AI into everyday business operations. Contract value, resulting subnet revenue and associated alpha buybacks were not disclosed; the filings do not independently audit each station or establish that every deployed model came from miners.
SN44 Score’s commercial platform has a large, named real-world rollout confirmed in a company release and an SEC filing. The rating reflects deployment and commercial relevance, with revenue and token-flow limits retained.
SN44 Score reports 143 free users and plans a Studio rebuild
SN44 Score says Score Studio has 143 individual free users, with 3% converting. The team explicitly excludes B2B partners paying for subnet outcomes from those figures, so they should not be used to measure the Avia deployment or total commercial adoption. Early feedback is driving a substantial product rebuild, followed by a planned relaunch across major platforms. This is a useful view of the consumer product’s early funnel and direction, not a relaunch already completed or a disclosed revenue amount.
SN44 Score supplies a concrete early user-conversion measurement and product decision, with the consumer and B2B businesses kept distinct.
Searches published headlines and summaries, not private bookmarks, unpublished leads or source-page contents. Duplicate headlines within an edition appear once; separate editions remain separate so follow-ups are preserved. Share the page URL to retain your filters.
