October 9, 2026 · Published daily
Subnet News — October 9, 2026
SN58 Attune opens its adaptive-robotics competition, SN105 Beam connects Hermes agents, and SN28 SayGM brings its gateway to the terminal. SN118 Ditto announces its first business client, SN54 Yanez reports a sponsor-bank partnership, and SN9 Iota shares fresh distributed-training results. Plus SN55 NIOME’s dated migration plan, SN30 Endure’s Forge collateral partnership, new tools and miner rules, and Const Reborn’s inference commentary—with the benchmark caveats included. Coverage runs through October 9 at 1:23 PM Eastern; later checks verify the sources within that window.
- Coverage window
- Oct 8, 2026, 12:31 PM – Oct 9, 2026, 1:23 PM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 27
Biggest stories first
Today's ranked news
Ledger partner alert for TAO holders: reseller-related loss reports under investigation
Ledger Support’s October 9 notice says it is investigating loss reports involving devices bought from CryptoBilis in Southeast Asia and has asked the reseller to pause sales and shipments. This matters to TAO holders using Ledger devices, including readers who use the site’s Ledger partner link. The notice concerns a specific reseller cohort; it does not establish a Bittensor compromise or a confirmed vulnerability in all Ledger devices. Ledger directs affected customers to its official support guidance. The cause and scope remain under investigation, so circulating loss totals and blanket claims that Ledger was hacked are not repeated here.
SN9 Iota reports SDK training across 139 consumer GPUs
Macrocosmos shared fresh SN9 Iota SDK results on October 8, one week after its Exploit Summit announcement. The team reports a 16B model trained across 139 RTX 4090/5090 GPUs with 1.8× faster wall time than its legacy stack, plus 30% shorter butterfly all-reduce times. It also reports 878 tokens per second serving gpt-oss-120b to 32 users on A6000s. A separate distributed 7B GRPO experiment improved MATH-500 from 63.6% to 68.4%. These are different workloads and team-reported measurements, not a single universal speedup or independent benchmark.
SN15 ORO opens AssistantEval’s replayable assistant tests
SN15 ORO pointed to AssistantEval on October 8. The public benchmark contains 34 tasks in resettable synthetic services, grading what happened in service logs rather than accepting an assistant’s claim that it finished. Its displayed results also expose reliability: 31% of assistant-task results that passed once failed on another attempt. This is useful scrutiny of agents handling inbox, calendar and other multi-step work. The site explicitly presents evaluation results, not demonstrated training gains or a guarantee about real accounts.
SN28 SayGM ships a CLI for model access and agent deployment
SN28 SayGM’s October 9 CLI announcement is backed by a public product page. It documents setup for Claude Code, Codex, Hermes and OpenClaw; terminal chat and shell-output prompts; and Hermes/OpenClaw deployment to Phala confidential VMs, SSH servers, EC2 or local Docker. The page specifies Node.js 22 or later and Windows, macOS and Linux support. This makes the gateway easier to use in existing coding and agent workflows. Documentation was checked; deployments and billing were not exercised.
SN30 Endure: Forge and TAO.com announce wrapped-alpha collateral collaboration
Forge announced on October 9 that it will build on TAO.com’s wstAlpha wrappers to bring alpha collateral to mainnet. Both TAO.com and SN30 Endure acknowledged the collaboration. The intended benefit is liquidity access without first selling the underlying alpha exposure. The language is explicitly forward-looking: these posts do not establish that borrowing markets are already open, disclose collateral parameters or independently verify the scope of any audit.
SN54 Yanez announces a US sponsor-bank partnership
SN54 Yanez announced a partnership with an unnamed US-based banking-as-a-service sponsor bank on October 8. Yuma highlighted the distribution opportunity on October 9: sponsor banks provide regulated infrastructure behind consumer fintech products. That potentially puts financial-crime tools closer to institutional workflows, but neither post names the bank or discloses contract value, rollout scope or revenue. This is a team-announced commercial development, not proof that all fintech customers of the bank have adopted SN54 Yanez.
SN55 NIOME sets October 28 for phased GENOME migration
SN55 NIOME’s October 9 announcement sets October 28 as the migration opening date. The first batch is intended to move 5% of GENOME’s snapshot market capitalization into SN55 NIOME, with later batches continuing the process. The team describes escrow, delivery verification and burning the original tokens only after successful delivery, with refunds if a swap fails. The portal link is promised when migration opens; this is not a swap already available today or an independently tested guarantee. At the 17:22 UTC market snapshot, SN55 NIOME was up 19.47% over 24 hours in alpha/TAO terms; that does not prove the announcement caused the move.
SN58 Attune opens Vector, an adaptive-robotics competition
Announced October 8 after the previous report’s cutoff, SN58 Attune is Robotensor’s competition for robots that adapt from demonstrations instead of needing a fresh model for every task. Its public site marks Vector live: submitted models face matched demonstrations, starting states and randomized simulation conditions. Horizon, combining human-video learning and world-action models, remains coming soon. New code rejects commitments made before a configured competition start block. This verifies the public competition launch and entry framework; it does not establish that miner rewards are already flowing.
SN105 Beam announces live Hermes Agent integration
SN105 Beam announced its Hermes Agent integration on October 9, bringing its data movement and streaming infrastructure into Nous Research’s agent environment. The practical point is letting an agent move data between cloud services, private infrastructure, applications and compute instead of leaving those steps to a human. This is a distinct integration from the previously covered OpenClaw skill. The announcement establishes the team’s live integration claim, not independently measured customer usage or support for every Beam transport feature.
SN118 Ditto announces its first business client and custom build
SN118 Ditto says it has secured its first business client less than two weeks after V2 launched. The proposed custom build connects the customer’s disconnected business applications, automates routine work and gives staff permissioned access through the mobile app. The customer is unnamed. Crucially, the description of a multi-six-figure company refers to the customer’s business, not a disclosed contract payment to SN118 Ditto. A commission-based partner program is still forthcoming; revenue-funded buybacks remain the team’s stated business objective, not a quantified new burn.
Protocol: spec 477 merge is confirmed; Finney activation is a separate check
The Subtensor repository merged the spec-477/SDK-11.3.1 update after the previous edition’s cutoff. That is verifiable code integration, not sufficient evidence that the finalized Finney runtime has activated it. Runtime 475 and Null Consensus discussions should likewise keep the mechanism’s opt-in behavior separate from deployment claims. This edition does not promote a merged runtime version into a completed network upgrade.
SN3 Teutonic changes how evaluation handles document boundaries
SN3 Teutonic published document-masked evaluation and token-weighted scoring on October 9. The new policy isolates document fragments, resets positions and excludes predictions across document boundaries. It also supports additional whole-document sampling from roughly 2K to 8K tokens. Deployment instructions require coordinated validator and evaluator policy identities, so this is a source-level evaluation change rather than proof that every current score already uses it.
SN14 Cacheon and SN46 Instant: Const highlights the inference race
Const’s October 9 post drew attention to an inference-speed comparison between SN14 Cacheon and SN46 Instant. The image reports 205.40 versus 162.61 tokens per second and lower first-reasoning latency for the shown SN14 Cacheon run. But the tests differ: SN14 Cacheon ran locally with GLM-5.3 NVFP4 on four B300s using SGLang; the SN46 Instant comparison used a GLM-5.3-Flash FP8 API run on two B300s with vLLM, according to the visible technical reply. Treat this as commentary on competitive optimization, not a controlled winner-takes-all benchmark. Separately, SN14 Cacheon reports its first customer, SN120 Affine, and a team-measured 40.9% throughput improvement for the optimized Qwen workload.
SN19 blockmachine announces live Arc RPC endpoints
SN19 blockmachine announced dedicated Arc mainnet RPC endpoints on October 8, with dollar-denominated pricing and payment in USDC. The team advertises pricing up to 60% below centralized alternatives, a comparison that depends on the actual endpoint plan and usage. The material development is support for another network through the subnet’s infrastructure product; no new customer total or revenue figure was disclosed.
SN22 Desearch announces a Raycast MCP Registry listing
SN22 Desearch says its search tools are now listed in Raycast’s MCP Registry, giving Raycast users another installation route for its search service. This October 9 announcement is separate from the October 6 Gemini CLI extension resurfacing in TAO Daily’s October 8 article. A registry listing reduces discovery and setup friction; it does not by itself demonstrate new paid usage.
SN25 UR records a deployment, with provider rollout still gated
SN25 UR’s primary launch document now records an October 8 deployment, contract calls and a validator activation block. It describes one launch validator and a weekly epoch configuration from epoch 1, effective at block 9,247,228. Operator enablement and reserved-slot upload still follow the epoch start. The record is more concrete than generic migration chatter, but it is the project’s deployment account: this edition did not independently replay every chain receipt or verify that all providers are operational and receiving rewards.
SN46 Instant’s public dashboard gets a Const spotlight
Const linked InstantNetwork on October 8 evening Eastern. Its public dashboard identifies SN46 Instant and exposes miner throughput, request failures and scoring. The inspected October 9 16:16 UTC epoch snapshot showed six miners and 164.8 tokens per second as the mean of miner medians, with 7 failures among 1,237 routed requests. These are workload- and epoch-specific operational observations, not audited revenue or a guarantee about a user’s own prompts. The pointer gives readers a way to inspect the network instead of relying only on screenshots.
SN49 Nepher demonstrates its robot policy on real hardware
SN49 Nepher’s October 9 demonstration shows an SO-ARM101 writing the team’s name and tracing its logo in simulation and on the physical arm. The team contrasts this with a July test where a circle that worked in simulation failed on hardware. This is a concrete sim-to-real progress demonstration for the shown task. It is not evidence of general-purpose manipulation, a broad benchmark win or readiness for every robot embodiment.
SN51 lium.io changes its missing-settlement fallback in code
SN51 lium.io merged a validator change so a missing settled window submits no new weights, retaining the weights already in force rather than sending a fallback burn vector. Inclusion reports are kept and retried instead of being lost if posting fails. This concerns a particular settlement failure path, not a blanket end to burns or a new revenue announcement. The merged implementation was reviewed; production adoption was not independently confirmed.
SN62 Ridges merges demand-sensitive upload pricing
SN62 Ridges merged competition-specific submission pricing. The documented rule starts at $5, increases about 15% with each purchased upload and halves every 30 minutes toward the floor. Payment burns alpha at the rate when the purchase lands, and unused burn is tracked against the coldkey for later uploads. Competition-bound tickets can expire unused without a refund. This is verified repository work that changes submission economics; live deployment and individual miner quotes were not independently exercised.
SN89 InfiniteQuant schedules a higher collateral floor for dust protection
SN89 InfiniteQuant’s source schedules the collateral minimum for Markets entity dust protection to rise from 120 to 6,100 alpha at October 10 00:00 UTC, which is October 9 at 8 PM Eastern. This applies to a qualifying active entity’s own coldkey stake on its hotkey. Dust protection helps avoid pruning while the entity earns no ordinary reward; it is not a share of the Markets reward pool or a universal deposit required of every token holder. The code’s approximate $5,000 valuation is dated, not a fixed dollar threshold.
SN100 Cortex opens mobile alpha access through TestFlight
SN100 Cortex announced on October 8 that its mobile application is available in alpha through TestFlight. This is an early-access distribution milestone, not a general App Store release or confirmation that every feature is production ready. It gives testers another way into the product while the team develops its wider agent tooling.
SN114 SOMA reports more inference integrations and traffic to SN97 Albedo
SN114 SOMA’s October 9 update names integrations with SN64 Chutes, SN28 SayGM and SN53 engy, and says it has begun proxying traffic to SN97 Albedo. The team frames this as a way to test its compression algorithms, reach prospective customers and improve inference efficiency. The SN64 Chutes connection was already covered; the broader provider list and SN97 Albedo traffic are the fresh context. No usage volume, measured savings or new revenue was disclosed in the post.
SN122 CookingTAO announces Claude Startups acceptance
SN122 CookingTAO says it has been accepted into Anthropic’s Claude Startups Program and already uses Claude tools in its mining-management portal. Program resources may support product development, but acceptance is not an Anthropic investment, endorsement of subnet economics or exclusive commercial partnership. A separate October 8 post targets public v2.2 access next week; that remains a planned release rather than a portal newly opened to everyone today.
Const catch-up: multisig safety, SN120 Affine and Gamma commentary
The required seven-day Const review also found relevant older comments. On October 4 he encouraged multisig use with QR-vault or Ledger signers; on October 7 he linked an SN120 Affine vision checkpoint and discussed its model identity. On October 2 he expressed optimism that Gamma would accelerate ecosystem development. These are dated operational advice, a model pointer and an opinion about a proposed direction—not new October 9 protocol activations. His short reactions to SN91 Cascade, SN41 Almanac and Taoswap were reviewed and deduplicated rather than turned into separate announcements.
SN17 404 shows Sundrift, another game-jam build
SN17 404 shared Sundrift on October 9, a mountain-pass drift racer built using its game generator with procedural 3JS assets. The point is another concrete example of the creation tool in use, rather than a new subnet launch. Const’s October 7 reaction to the earlier Farseek game is retained as dated commentary in the source review, not relabeled as today’s product release.
SN66 Conjectures raises proof-submission limits in merged code
SN66 Conjectures merged a change allowing proofs up to five million tokens and 30 MiB. That widens the space for long formal proofs while leaving successful verification and reward eligibility as separate gates. A circulating first-contributor payout report was also reviewed, but the payout amount and treasury execution were not independently checked on-chain in this edition.
SN90 KubeTEE documents Nemotron Omni’s 256K context
SN90 KubeTEE’s repository update documents a 256K context configuration verified with NIM 2.0.13. This is a specific inference-serving configuration update, useful to operators planning longer inputs. It does not establish that every hosted endpoint has adopted the new configuration or that every long-context workload will have the same performance. The repository describes the endpoint as evaluation-only pending a commercial license grant; it should not be presented as a generally billed production service.
SN95 Actual adds an in-browser toks playground
SN95 Actual linked a browser playground for toks on October 9. The repository explains that it runs the portable C core compiled to WebAssembly and produces the same token IDs as the native library. Browser timings use the scalar tier, so they should not be compared directly with the native assembly-kernel speed claims. The Apache 2.0 licensing change was covered yesterday; today’s distinct development is a simpler way to inspect tokenization.
SN116 Carbon’s large market move is not proof of a production launch
SN116 Carbon triggered review with a 23.55% 24-hour and 2,078.43% seven-day rise in alpha/TAO at the October 9 17:22 UTC snapshot. The project’s public status document describes simulation and research across battery cold plates, AI-chip cooling, motors and photonics, while distinguishing experiments from customer qualification and production reward-network readiness. Fresh repository activity does not resolve those gates. The business thesis is physics optimization; a verified new production launch or commercial contract explaining this price move was not found.
Audit trail
Coverage notes
Searched
- All 128 current subnet profiles and dated public SubnetRadar briefs; 187 associated repository paths.
- 201 associated public X account samples, including teams and partners; signed-in @ShizzyUnchained direct source checks.
- Signed-in Grok broad discovery and focused omission review, explicitly including @const_reborn.
- Signed-in AlphaGap dashboard; SubnetRadar news and Moments; TAO Daily, TAO Media and primary product pages.
- Previous seven editions, market outliers, emission leaders and unresolved incidents.
Public-data gaps
- Associated account samples and Grok are discovery aids, not exhaustive X search for every SN number and alias. Private Discord discussions were not independently read; unverified allegations and unconfirmed activation claims were withheld.
- Repository commits prove published code, not live validator adoption. Team benchmarks and undisclosed commercial deals are attributed; they are not independently audited results.