Subnet teams' own channels, X posts, press, newsletters and events read from the chain — GitHub lives in its own tab
REAX produced the world's top-ranked decision model in weeks, using open miner competition and model blending to achieve a score of 74.0. [...] The post Bittensor Produces the World’s Top-Ranked Decision Model With REAX 1.3 appeared first on The TAO Daily.
The TAO Dailysource ↗
Openτensor Foundaτion@opentensorBittensor’s open competition just produced a world-leading AI decision model. REAX 1.3 has the highest score among 270+ entries on the Decision Index hosted on @huggingface, ahead of leading models from Perplexity and Fastino.
Quoting @reaxlabs
🥇 REAX 1.3 is #1 on the Hugging Face Decision Index. We produced the best decision model in the world. 274 models. One top spot. Built on Bittensor $TAO. Four weeks ago, Subnet 92 didn't exist. Then the miners went to work: an open competition, two selected miner models, one post-production run. Now REAX 1.3 sits at the top of Decision Index 0.4.1 with a Full score of 74.0. This is what decentralized AI looks like when it ships.⚡ Live today on a Jev-compatible API decisionmodels.io huggingface.co/spaces/multimodalart/jev-decisio…
Taosis News: the week’s articles from the subnet teams, the press and community newsletters, every Friday. Taosis Weekly: the week’s movers, best yields and chain events, with dollar values, every Monday. Past issues
REAX@reaxlabsThanks for the excellent collaboration, @trishoolai! AI security and decision models are a match made in heaven. REAX 🤝 TRISHOOL
Quoting @trishoolai
Decentralised AI, now with an EU address. 🇪🇺 Trishool (SN23) and REAX (SN92) are partnering to serve the EU market. REAX is the team behind @decisionmodels. Their decision models now power EU Mode, the AI security product from the team behind Trishool. Why an EU partner? Many European companies cannot send sensitive data outside Europe. The AI that checks their AI must follow the same rule. A security layer that sends prompts to other regions creates the risk it should remove. REAX runs its models on GPU servers in EU data centres. That is why we chose them. This is the Bittensor economy at work. Subnets that supply each other, build with each other and ship together. SN23 🤝 SN92
REAX@reaxlabsREAX 1.3 is open weights, Apache-2.0 license. huggingface.co/decisionmodels/REAX-1.3
Linkhuggingface.cohuggingface.co/decisionmodels/REAX-1.3REAX@reaxlabs🥇 REAX 1.3 is #1 on the Hugging Face Decision Index. We produced the best decision model in the world. 274 models. One top spot. Built on Bittensor $TAO. Four weeks ago, Subnet 92 didn't exist. Then the miners went to work: an open competition, two selected miner models, one post-production run. Now REAX 1.3 sits at the top of Decision Index 0.4.1 with a Full score of 74.0. This is what decentralized AI looks like when it ships.⚡ Live today on a Jev-compatible API decisionmodels.io huggingface.co/spaces/multimodalart/jev-decisio…
Linkdecisionmodels.iodecisionmodels.ioShare of the subnet's miner emission routed to burn: 29% → 57%.
Detected 4d agoView subnet →
On Taosis
Swap TAO ⇄ USDC
USDC on Ethereum to TAO, or back. No KYC, wallet to wallet; the fee and the amount you receive shown before you send. Not instant: run by Taosis, paid by hand within hours.
REAX@reaxlabsOfficial scores: We'll submit REAX 1.3 to the Hugging Face Decision Index later today, so results are pending. According to our own measurements we see a real chance at the top 10 of the index, maybe even #1, but let's wait for the official numbers before declaring victory. We are the maintainers of JevBench, so REAX 1.3 gets no row there.
REAX@reaxlabsAPI: decisionmodels.io/models Engine (MIT): github.com/reaxlabs/reax-engine Open Weights (Apache 2.0): huggingface.co/decisionmodels/REAX-1.3 Thank you to every miner on Subnet 92 who competed.
Linkdecisionmodels.iodecisionmodels.io/modelsREAX@reaxlabsIntroducing REAX 1.3 Our new open decision model, built with two models from our Bittensor miner competition and strong open models. Results: our own run on public Hugging Face Decision Index items. Available via API at decisionmodels.io
Linkdecisionmodels.iodecisionmodels.ioREAX@reaxlabsHow it was built: - Miners submitted 98 entries to our model training competition - Our post-production picked two and blended them into top open decision models: PelaAI/KnowLine-31B and vllm-sr/d3-ultra - Both parts run in parallel on one RTX PRO 6000 and their probabilities are combined into one answer - Served in FP8 by our open-source REAX Engine (MIT) - The miner fine-tunes derive from tattrongvu/Jeffan-31b-1710 (Gemma-4-31B line) and baselabs/mammoth-1 (Qwen3.5-27B line)
Share of the subnet's miner emission routed to burn: 3% → 29%.
Detected 7d agoView subnet →
REAX announces a smaller inference allocation from November 2 and changes when champions' protected payment period starts. Validator execution is unverified.
Tao Outsidersource ↗
REAX@reaxlabsSN91, Cascade: "We connected Ephemeris, our time-series forecasting API, with Decision Models (powered by @reaxlabs) to build a system that tackles this. Ephemeris forecasts how market conditions may develop."
Quoting @cascade_sn91
REAX SN92 records a 230.55-alpha lock as miners prepare for a conditional November 2 step tying paid inference capacity to customer demand.
Tao Outsidersource ↗
Share of the subnet's miner emission routed to burn: 37% → 3%.
Detected 10d agoView subnet →
Share of the subnet's miner emission routed to burn: 0% → 37%.
Detected 10d agoView subnet →
REAX@reaxlabsSN92: the next phase Subnets that last are the ones where real revenue pays the bills. That's the direction I laid out on 18 Oct in the subnet discord, and after Friday's crowning slots SN92 takes the next step. Phase 1: Build (done). The launch phase was deliberately generous: test the network at scale, bring many miners in, learn fast. It delivered: •engine techniques that make our models much faster and cheaper to run, now going into decisionmodels.io •real know-how on training better decision models; the strongest entries that clearly beats their base go to the Hugging Face Decision Index •a community of miners who know this subnet inside out Phase 2: Earn (now). Spending follows real revenue and what operations actually need: •Both competition tracks pause after Friday's slots. •Inference rewards move step by step toward real customer demand, starting next week, each step announced here first. •Fair to everyone: every crowning is paid properly (full share for 3 days, then down to zero by day 7, then keep-alive, so no champion gets pruned), and everything already earned stays paid. Phase 3: Scale. As paying customers grow, rewards grow with them, paid by real demand instead of emission. That's escape velocity. Why this works •Speed: miners in the engine competition made our models up to 3–4× faster on the same GPUs, verified against strict correctness checks. That goes straight into production. •Cost: decision models are cheap to serve. A single RTX PRO 6000 handles 30 to 95 decisions per second in our measurements, while large LLMs need clusters of the most expensive GPUs for a handful of users. This is the type of business that can pay for itself. •Demand: weekly requests to decision models on OpenRouter doubled in four weeks, to 1.67 billion a week (source: openrouter.ai/rankings). •Traction: decisionmodels.io is growing organically. One customer who found us on their own sent 1.7 million requests in a single week. We are well…
Linkdecisionmodels.iodecisionmodels.ioAffine isolates dataset batches, TorchLean splits its competition across six boards, and REAX adds an October 30 Generalist slot with a new cutoff.
Tao Outsidersource ↗
REAX 1.2 from Bittensor SN92 releases Apache-2.0 weights and uses abstention more often. Read both accuracy denominators from the team's internal tests.
Tao Outsidersource ↗
Validator hotkey 5DM7CPqPKtMSADhFKYsstsCS4Tm4Kd6PMXoh6DdqY4MtxmtX announced new child keys on SN92, SN95, applying at block 9,362,340 (~24h). Stake delegated to it there routes through its children after that. It replaces the changes that were pending on SN92 (block 9,356,624) and SN95 (block 9,356,623), which will not apply then.
Between blocks 9,355,112 and 9,355,177
REAX@reaxlabsREAX 1.2 is out as open weights under Apache-2.0, a fine-tune of REAX 1. huggingface.co/decisionmodels/reax-1.2 REAX 1.1 is up too: huggingface.co/decisionmodels/reax-1.1
Linkhuggingface.cohuggingface.co/decisionmodels/reax-1.2Validator hotkey 5DM7CPqPKtMSADhFKYsstsCS4Tm4Kd6PMXoh6DdqY4MtxmtX announced new child keys on SN92, SN95, applying at block 9,356,623 (~24h). Stake delegated to it there routes through its children after that.
Between blocks 9,349,373 and 9,349,433