REAX.
Menu

REAX / COMPETITIONS

Competitions

Besides serving inference, miners can compete to improve the network itself: a faster engine, and specialist decision models for single use cases.

Use-case tracks

Fine-tune a small open-weight decision model for one use case. Each track has its own page with a real example item, the data, step-by-step submission and the live leaderboard.

Where miner emissions go

  • 50% Inference miners, distributed by contribution (hosting the pinned decision models, scored on correct answers and speed).
  • 20% REAX Engine competition (permanent).
  • 15% Use-case fine-tune track A (rotates every 2 weeks).
  • 15% Use-case fine-tune track B (rotates every 2 weeks).

Both competitions open on 16 October 2026, 12:00 UTC. Before that time, all miner emission goes to inference miners. Nothing is burned: a competition share that nobody wins, or that has decayed, flows back to the inference miners. The split was approved by the subnet owner. It can change later; any change is announced in advance on reax.co and mine.decisionmodels.io, with the date it takes effect. The full picture is on How mining works.

Overview

Besides serving inference, miners on subnet 92 can compete to improve the network itself. Two kinds of competition run side by side:

  1. Engine competition, "Make the REAX Engine faster". Same model, same answers, measurably faster.
  2. Use-case tracks. Fine-tune a small open-weight decision model for one use case. Two tracks run at the same time; the first two are demand_forecasting and risk_assessment.

Both are judged by validators on their own hardware, on fresh items nobody has seen. Release v1.2.0-rc1 contains the competition code.

Winner takes most

Each competition works winner-takes-most with decay, similar to Affine (subnet 120): the best valid submission takes most of the competition's share, and the share decays over time. To keep noise and copies from paying, a challenger needs a required margin over the current champion and a confirmation period: it has to win again on a second, fresh evaluation before the title changes. Ties go to the earlier commitment, and a copy of the champion does not pay.

Rules can change; any change is announced in advance with the date it takes effect. Rewards are a share of miner emissions, set by validators from benchmark results. They are a protocol mechanism, not a promise of any amount, and can be zero.

OPENS 16 OCT 2026 · 12:00 UTC

Make the REAX Engine faster

Status: Submissions and measurements open 16 Oct 12:00 UTC. Engines are measured on a dedicated RTX PRO 6000 (96 GB) machine that holds no keys, against the reference in the same session.

Permanent track, 20% of miner emission. Serve the existing s1-fast model (Plumb-4B, pinned weights) faster than the reference engine. Same answers, less time per decision. A winning engine can become a drop-in runtime for all serving miners.

What you may and may not change

  • You may change serving code: kernels, batching, scheduling, CUDA graphs, memory layout, tokenisation, process layout. Quantisation only if the answers stay inside the correctness gate.
  • You may not change the weights, the prompt, the readout, or anything that changes answers beyond the gate. No caching across requests. No network access.

How validators measure

  • Relative to the reference. Every validator measures every new submission on its own GPU, against the reference engine in the same session, interleaved. Only the relative speed counts, so a bigger or smaller GPU does not matter. Our validator runs on an RTX PRO 6000 96 GB.
  • Fresh items. Validators generate fresh items after your engine has started, with a secret seed. Items never repeat, so a lookup table or cache cannot help.
  • Correctness gate. Every timed answer is compared with the reference answer. Mean logit distance must be at most 0.10, every item at most 1.0, and the top choice must agree on at least 99% of items.
  • Anti-cache probes. They repeat a state with a different question, and a question with a different state. A cache keyed on part of the input fails.
  • Determinism and latency. A determinism re-run and a latency check (p50 at most 1.25 times the reference) must also pass.
  • Sealed run. The engine runs with no network, read-only files and resource caps. Nothing the engine reports is trusted; validators use their own clock and recompute every answer.

Who wins

  • A submission counts only if it passes every gate and is at least 3% faster than the reference.
  • A challenger takes the champion's title only with at least 3% more than the champion, and only after a second measurement on fresh items, in a later cycle at least 6 hours later, also clears the bar. Ties go to the earlier on-chain commitment.
  • The champion is re-measured whenever it is challenged. If the re-measurement fails, the title is vacated.
  • Share over time. When crowned, the champion gets the whole engine share (20% of miner emission). That share then halves linearly over 14 days (75% after 7 days, 50% from day 14 on); the decayed part goes back to inference. A new champion starts at 100% again.

Leaderboard

Coming soon — the leaderboard appears here when the competition opens.

OPENS 16 OCT 2026 · 12:00 UTC

Specialist decision models per use case

Two use-case tracks run at a time, each with 15% of miner emission. The goal is an open-weight Jev-compatible decision model that beats the current best model on one use case. Validators score the incumbent and the challengers on fresh sealed items that nobody has seen.

Tracks

demand_forecasting and risk_assessment, both from 16 October 12:00 UTC to 30 October 12:00 UTC. Tracks then rotate every 14 days; the next tracks are announced before they open. A commitment for a track that is not active is ignored. Until a track is active, its share stays with inference.

What you submit

  • Weights only, no code. A Hugging Face model repo at a pinned 40-character revision, safetensors only, no trust_remote_code.
  • Size: at most 9 B parameters and at most 20 GiB of weights.
  • Licence that allows us to host and serve the model: Apache-2.0, MIT, BSD-2-Clause or BSD-3-Clause, with a LICENSE file. Other licences only with the subnet owner's written approval.
  • Architecture: Qwen3, Qwen3.5, Gemma 3 or Llama. Gemma 4 checkpoints are not accepted in this release.
  • Validators run your model on their own pinned serving setup with a fixed prompt and readout, so no miner code runs.

How validators score

  • Each round, the incumbent and every challenger answer the same fresh draw from a sealed pool of items.
  • Score = 0.75 x competence + 0.25 x calibration. Competence is the chance-corrected correct-answer score. Calibration rewards honest probabilities.

Who wins

  • A challenger must beat the incumbent by max(1.5 points, 2 standard errors).
  • It is then confirmed on a second, disjoint fresh draw at least 12 hours later. Earlier commitments are processed first.
  • Entry floor: a submission must beat the s1-fast model (Plumb-4B, the model miners already serve) on the sealed items by at least 1.5 points: demand forecasting 24.8 (s1-fast scores 23.3), risk assessment 33.7 (s1-fast scores 32.2). Below the floor a submission is not ranked; the track's share then goes to inference miners.
  • Inside a track: rank 1 gets 75%, rank 2 17%, rank 3 8%. Ranks 2 and 3 are the most recent former champions that are still committed.
  • Winning models. Winning specialists may be served on decisionmodels.io.

Leaderboard

Coming soon — the leaderboard appears here when the competition opens.

Rewards

Competitions together take 50% of miner emission: engine 20%, the two use-case tracks 15% + 15%. Inference keeps 50%. If an engine has no champion, or a track has no valid result or a missing rank, that share goes back to inference. Nothing is burned.

Timeline

Competition timeline
WhenWhat
Until 16 Oct 2026, 12:00 UTC100% inference
16 Oct 2026, 12:00 UTCEngine competition and the two first use-case tracks open
Every 6 hoursValidators run an evaluation cycle for new engine submissions
About every 12 hoursUse-case rounds; a new leader needs a confirmation at least 12 h later
30 Oct 2026, 12:00 UTCThe first two tracks end; the next tracks follow every 14 days
14 days after crowningAn engine champion's share has halved

Leaderboards appear on this page and on the track pages when the competitions open.

How to submit

You need a hotkey registered on subnet 92. A competition hotkey does not serve traffic; if you want to serve and compete, register two UIDs.

Engine

  1. Build a container image that implements the engine contract (a unix socket, POST /internal/v1/decide, GET /health). Add the label ai.reax.engine.hotkey=<your hotkey ss58>. Push it to ghcr.io, docker.io or quay.io.
  2. Check locally with the validator measurement: reax-engine-eval --reference <ref> --candidate <image> --weights <model dir>
  3. Commit the image by digest: reax-engine-submit <image>@sha256:<digest>. On chain this is reax-eng/1:s1-fast:<image>@sha256:<64 hex>. One commitment per hotkey; the digest cannot change afterwards.

Use-case track

  1. Fine-tune and upload your model to Hugging Face (safetensors, LICENSE, tokenizer, reax-specialist.json manifest, at most 9 B). The tokenizer must carry the chat format of the model family: Qwen ChatML, Gemma 3 or Llama 3 special tokens. Gemma 4 checkpoints are not accepted in this release.
  2. Get the public train and dev data (CC-BY-4.0; the sealed split stays private): huggingface-cli download decisionmodels/sn92-specialists --repo-type dataset --local-dir sn92-specialists (dataset page).
  3. Check locally on the public dev split: reax-spec-eval ./my-model --dev ./sn92-specialists --category demand_forecasting --hotkey <your hotkey>. The number is a preview, not a ranking. It needs Docker with the NVIDIA runtime and a GPU with about 25 GB free.
  4. Commit on chain: reax-spec/1:<track_id>:<hf_repo>@<40-hex revision>. You may commit while the repo is still private; make it public within 7,200 blocks (about 24 hours) and keep it public. Limits per coldkey: 2 open submissions per track, 4 new submissions per 7 days.

Anti-copy

  • Engine: the image label binds a digest to the hotkey that committed it, so nobody can claim someone else's digest. A rebuilt copy must still beat the champion by 3%, and a lucky measurement needs a second one to confirm it.
  • Use-case: copies are detected by a tensor-level hash (re-sharding or renaming files does not hide a copy) and by comparing weight changes against the declared base model. A derived submission must beat its parent head-to-head. Earlier commitments are processed first.
  • Items are fresh and sealed, so there is no answer key to copy.

FAQ

Can I serve and compete with one hotkey?

No. A competition hotkey does not serve traffic. Register a second UID if you want both.

Does my GPU size matter for the engine competition?

No. Validators compare your engine with the reference on the same GPU in the same session. Only the relative speed counts.

Can my engine use a different model or lower precision?

Only if every answer stays within the correctness gate. Different weights, cached answers, or shortcuts that change answers fail the check.

What happens if nobody wins?

The unclaimed share goes back to inference. Nothing is burned.

Are the competition rules fixed?

No. They can change, and any change is announced in advance on reax.co and mine.decisionmodels.io with the date it takes effect.

What happens to winning use-case models?

Winning specialists may be served on decisionmodels.io. Models must have a licence that allows us to host them.

What happens after 30 October?

The next pair of tracks is announced before it opens. Until a track is active, its share stays with inference.

Live serving scores are on the miner leaderboard. Want to mine? See How to mine.