How mining works

01 / WHAT MINERS DO

Mining is hosting.

A REAX miner runs one of the pinned decision models on its own GPU and answers the questions that validators send. Nobody brings their own model for this: the exact weights and the runtime image are fixed, so every miner serves the same thing and answers can be checked.

1. Run the pinned model. Start the published miner software on a machine with a supported GPU and a public TCP port.
2. Answer validator rounds. About every 2 minutes our validator sends your miner fresh decision items and waits for answers.
3. Get scored. Correct answers count first, speed adds on top. The score decides your weight, and the chain shares the miner emission by weight.

Rewards are paid in the subnet's own token, are variable and can be zero. Hardware, electricity and the registration fee are your costs. Nothing on this page is investment advice.

02 / POOLS

Pools: four kinds of machine, four pools.

A pool is one model on one kind of hardware. Release 1.2.0-rc1 has four pools, from an 80 GB datacenter GPU down to a plain CPU. One hotkey serves one pool. Each pool is scored on its own.

• Tiers. The inference part of miner emission is split by tier: GPU 80%, small GPU 15%, CPU 5%. Inside the GPU tier, s1-fast gets 35% and s1-pro 65%. The average per miner in a tier is at most half of the tier above.
• Empty pools. If a pool has no miner, its share passes to the other pools. Nothing is burned.
• Release constants. Tiers, shares and targets change only with a new release, announced in advance.
• Not yet calibrated on real hardware: s1-lite on small GPUs and Jetson devices. The pool is open, and measured speeds will be published.

Only the exact pinned weights in the pinned runtime image count: bf16 for s1-fast and s1-pro, the pinned 8-bit GGUF file for s1-lite and s1-cpu. Any other quantised build (FP8, INT4, a different GGUF) scores badly or is struck. The "share of the inference part" is 100% of miner emission before the competitions open and 50% after.

03 / SCORING

How a miner is scored.

• Fresh questions every round. Our validator (UID 3) asks every miner new decision items about every 2 minutes and compares the answers with its own reference run of the same pinned model. Items are fresh every round, so there is nothing to cache.
• Correct answers first. Answers that match the reference score. Answers that drift too far score zero for the round.
• Speed adds up to 20% of a round's score. Only a real speed-up helps, measured on the validator's own clock.
• A moving average. Round scores feed an exponential moving average (EMA). A new miner starts at zero and builds up over about an hour of good rounds.
• Downtime or the wrong model scores zero. A miner that is offline, or serves a different or quantised model, earns nothing in that round.

The validator now runs on an RTX PRO 6000 with 96 GB of memory. Live scores for every miner are on the miner leaderboard. This is our validator's view; other validators may score differently.

Capacity-based scoring (release 1.2.0-rc1). Inside a pool, a miner's weight is its score times its verified capacity, capped at the pool's target. Capacity is measured by surprise load probes (about one round in ten, at a time you cannot predict) that hit all UIDs of a pool at the same moment. A UID that was never measured has capacity 0 until its first probe, which comes in the next round it answers.

04 / THE SPLIT

Where miner emissions go.

Half of the miner emission goes to inference miners, half to competitions. The split is decided and takes effect when competitions open on 16 October 2026, 12:00 UTC. Until then, 100% goes to inference miners.

• 50% Inference. Distributed by contribution: hosting the pinned models, scored as above.
• 20% REAX Engine competition. A permanent track: make the engine that serves the model faster.
• 15% Use-case fine-tune track A. Rotates every 2 weeks.
• 15% Use-case fine-tune track B. Rotates every 2 weeks.

• Winner takes most, with decay. In each competition the best valid submission gets most of that competition's share, and the share decays over time (similar to Affine, subnet 120).
• Nothing is burned. If a competition has no winner, or part of its share decays, that part flows back to the inference miners.
• Until 16 October 2026, 12:00 UTC: all miner emission goes to inference miners.

The split can change later. Any change is announced in advance on reax.co and mine.decisionmodels.io, with the date it takes effect.

05 / COMPETITIONS

Three competitions, one network.

The rules, the evaluation tools and the leaderboards are on the competitions page.

06 / SEVERAL NODES

More than one machine.

You can run several nodes. Each node needs its own hotkey and UID, and serves one pool. Weight is score times verified capacity, so many hotkeys on one machine share that machine's throughput and add nothing, while one hotkey with more GPUs behind it counts more, up to the pool's target. A bad answer is a strike and resets the moving average and the capacity.

07 / START

Ready to set up a node?

The setup guide is written for you and for your coding agent: hardware check, rented GPUs, wallet, registration and keeping the node running.

Set up a miner ↗See the live leaderboardCompetitions

Miner rewards are variable and can be zero.