REAX.
Menu
THE REAX NETWORK Live · netuid 92

AI by the people,
for the people.

REAX is Bittensor subnet 92: an open inference network for decision models. Anyone with a capable GPU can host the models, validators check every answer, and the network rewards correct, fast serving.

Use the API Mine Read the code

01 / HOSTING IS MINING

A decision model is an AI model that answers one narrow question fast: route this request, escalate this case, choose the next action.

Open models,
hosted by many.

Miners run the decision models on their own GPUs. Validators check every answer against their own copy of the same model and time it. Correct, fast miners get more weight.

  1. 01 · SERVE

    Miners host the models

    Each miner runs the same pinned model and runtime, so every answer can be checked.

  2. 02 · VERIFY

    Validators check

    Fresh questions every round. Wrong answers or swapped models score zero.

  3. 03 · REWARD

    Measurements earn weight

    Correct answers first, speed second.

  4. 04 · IMPROVE

    The network gets better

    The generalist track for better decision models is open since 21 October (base: REAX 1). The engine track for a faster engine is open since 20 October.

PART 1 · INFERENCE · LIVE

Host the decision models

Four pools: s1-fast and s1-pro on NVIDIA GPUs, s1-lite on small GPUs incl. Jetson Orin, s1-cpu on plain CPUs.

PART 2 · COMPETITIONS · GENERALIST + ENGINE · OPEN

Make it faster, make it smarter

The generalist track is open since 21 October 2026, 13:00 UTC; its base model is REAX 1. The engine track is open since 20 October 2026, 18:00 UTC.

How mining works

02 / THE FIRST PRODUCT

Customers are served
on decisionmodels.io.

Decision Models is the customer-facing product: a Jev-compatible API for open decision models, billed in euros. Customers sign up and pay there, not here.

Decision Modelss1-fast, s1-pro, s1-vision and Quyet-1.0-Large behind one API. Peer-to-Peer keys can opt in to network processing: s1-fast requests may then be answered by REAX miners with a passing track record, each answer checked against our own GPU. EU-tier requests never go to the network.LIVE · DECISIONMODELS.IO

03 / WHY DECISION MODELS

Millions of small decisions.
Most don't need a giant model.

AI agents route requests, choose next actions and decide when to escalate. Each needs a fast, calibrated answer to one question. Decision models are fast, cheap and smart.

Route this request?Escalate this case?Choose the next action?

04 / GET STARTED

Use it. Mine it.
Check it.

USE

Call decision models

One Jev-compatible API with an API key. Billing is in euros, nothing on chain.

Use the API

MINE

Serve on your GPU

A setup guide for you and your coding agent: hardware, wallet, registration, running the node.

Start mining

VALIDATE

Run a validator

Recompute miners' answers, measure speed, set weights. The code is public.

Validator guide

REAX ENGINE

REAX Engine.
The code is public.

A prefill-only inference engine for decision models: one forward pass, no token generation. REAX Engine License 1.1 (MIT text plus one condition): the code is public and free for individuals and small companies; larger organisations need a separate written licence. Mining and validating on subnet 92 are always free.

≈3.1×decisions per second vs. the reference server (Decision-Index-style traffic)
2.6×decisions per second on the JevBench public set
Samebenchmark scores, within noise

Measured with Perplexity Decider v1.1 27B in FP8 on one RTX PRO 6000 Blackwell GPU, against the model's reference server with the same harness.

Experimental shared-state mode: 6.0× (31.73 to 31.98 vs. 5.24 to 5.28 decisions per second) on Decision-Index-style traffic where requests carry several questions about one state; single-question requests stay at about 3.1×. No measurable change in the Decision Index. It is an approximation: 2.1 % of multi-question top-1 answers differ.

github.com/reaxlabs/reax-engine

05 / REWARDS

Where miner
emissions go.

  • 50% Inference miners, shared by contribution (correct answers, then speed).
  • 20% REAX Engine competition: a faster engine for s1-fast (permanent).
  • 30% Generalist decision-model track, open since 21 October 2026, 13:00 UTC; inside it rank 1 gets 75%, rank 2 17% and rank 3 8%.

The generalist track is open since 21 October 2026, 13:00 UTC; the engine track is open since 20 October 2026, 18:00 UTC. The part of the 20 % engine share without an approved award goes to inference miners; nothing is burned. The split can change; changes are announced in advance.

The two use-case tracks (demand forecasting and risk assessment) ran from 16 October 2026, 12:00 UTC and ended when the generalist track opened; for 7 days their champions and approved former champions keep a bounded keep-alive as pure entrants.

Rewards are variable and can be zero; hardware and registration are your costs. Nothing on this site is investment advice.

Scoring in detail   Competitions

06 / STATUS

Live on mainnet.
Netuid 92.

OPEN NETWORK · LIVE

Anyone can mine.

Since
14 October 2026 on Bittensor mainnet.
Registration
Open. The chain sets the burn cost; check it before you register.
Pools
Four: s1-fast, s1-pro, s1-lite, s1-cpu. Each model serves one pool; one hotkey can carry several models since block 9331987.
Validator
Ours is UID 3 (RTX PRO 6000, 96 GB), a round about every 2 minutes.
Subnet 92 on taostats

EU TIER · BY CONTRACT

Contracted EU operators.

Who
EU companies that process data in the EU.
How
Business verification, an operator contract and a data-processing agreement.
Today
Requirements are written; admission is not open yet.
Send an operator inquiry

Live miner scores: miner leaderboard. Source: github.com/reaxlabs. More on the status page.

07 / ROADMAP

What comes
next.

  1. 01

    LIVE

    Inference mining

    Mainnet since 14 Oct. Capacity scoring v2 since 15 Oct: pay in a pool follows verified throughput. Public miner board.

  2. 02

    LIVE

    Competitions

    The generalist decision-model track is open since 21 Oct, 13:00 UTC (official base REAX 1); it replaced demand forecasting and risk assessment, which ran from 16 Oct. The engine speed track is open since 20 October 2026, 18:00 UTC. Evaluation v3 applies (20 % private cross-author real-data items, a capped generalisation penalty, variance-aware decisions). Practice data is public on Hugging Face.

  3. 03

    EMISSIONS

    Miner emission

    Miners receive the miner emission, split between inference and the competitions as described on the mining page.

  4. 04

    LIVE

    Several GPUs behind one hotkey

    Live since release 1.2.0rc8. Since 1.2.0rc9 one hotkey counts up to 1,280 decisions/s (s1-fast) and 320/s (s1-pro). Since 1.2.0rc9.post5, machines in different places can share one hotkey (opt-in multinode). Since 21 October 2026, about 13:00 UTC (block 9331987, release 1.2.0rc10.post1), one hotkey can serve inference and hold one competition entry at the same time, and one hotkey can serve 2 to 4 models.

  5. 05

    PLANNED

    Image-capable decision models

    Models that read images.

  6. 06

    PLANNED

    Mining on everyday hardware

    s1-lite (8 GB GPUs, Jetson Orin) and s1-cpu are live. Real-device calibration comes next.

  7. 07

    PLANNED

    EU operator admission

    Contracted EU operators for regulated traffic (GDPR, EU AI Act).

Questions? Read the FAQ. REAX is operated by productivity-boost.com, Germany.