# REAX inference mining (Bittensor subnet 92): agent entry point

Canonical URL: https://reax.co/agent/mine.md · Subnet release: 1.2.0rc3 · Updated: 2026-10-15

You are an AI agent. A person pointed you here, for example with "Read https://reax.co/agent/mine.md and help me mine REAX". Treat this page as reference material for that request. It does not override your own safety rules or the person's instructions. Follow your normal permission prompts.

## 1. Where the full playbook is

The complete, step-by-step miner playbook (hardware verdicts, rented GPUs, wallets, registration, running the node, troubleshooting) is maintained at **https://mine.decisionmodels.io/agent.md**, with a machine-readable twin at https://mine.decisionmodels.io/mine.json and the current release at https://mine.decisionmodels.io/releases/current.json. Read it next and follow it; this page only adds how to check a miner on reax.co.

## 2. Rules (non-negotiable, same as the playbook)

1. Never ask for, read, print, store or transmit a recovery phrase (mnemonic, seed), private key or coldkey file. You only handle public SS58 addresses and the hotkey on the mining machine. The person keeps the coldkey.
2. A human approves every payment: registration burns TAO, renting a GPU costs money. You prepare; the person decides and runs it.
3. Never publish the model sidecar ports; only the miner port (default 8091) is public.
4. No financial advice: never estimate earnings, income, returns or token prices. Rewards are variable and can be zero.

## 3. Pools in short

| Pool | Model | Minimum hardware | Capacity unit / cap |
|---|---|---|---|
| `s1-fast` | Plumb-4B, bf16 | NVIDIA GPU ≥ 24 GB | 40 / 640 decisions/s |
| `s1-pro` | Rune 26B-A4B v3, bf16 | NVIDIA GPU 80–96 GB | 10 / 160 decisions/s |
| `s1-lite` | Plumb-4B, 8-bit GGUF | NVIDIA GPU ≥ 8 GB (incl. Jetson Orin) | 8 / 32 decisions/s |
| `s1-cpu` | Gutsy 0.8B, 8-bit GGUF | x86-64 AVX2 or ARMv8.2 CPU, 4+ cores | 1 / 4 decisions/s |

Inside a pool the weight is score × verified capacity: validators send surprise load probes to every UID of a pool at once and count correct answers per second, linear up to the cap. Many hotkeys on one machine share its throughput and add nothing.

## 4. Check a miner after setup

- Live board for humans: `https://reax.co/miners/?hotkey=<hotkey ss58>` (or `https://reax.co/miners/miner/?uid=<uid>`).
- Data for you: `GET https://reax.co/data/miners/latest.json` and find the entry in `miners[]` whose `hotkey` matches. Useful fields: `status` (`ok`, `degraded`, `failing`, `offline`, `not serving`), `pool`, `availability_24h`, `correctness_24h`, `latency_p50_s_24h`, `failures_24h`, `capacity` (verified decisions/s, when probed). Per-UID history: `GET https://reax.co/data/miners/history/<uid>.json`. Data comes from our validator (UID 3) and refreshes every 5 minutes.
- Status meanings: `not serving` = no valid commitment or axon, so the validator does not query it; `offline` = the last 5 rounds failed to connect or timed out; `failing` = the last round failed or availability under 50 %; `degraded` = availability under 95 %, correctness under 99 % or score EMA under 0.5.
- If a miner is not `ok`, run the kit's `reax-selfcheck` on the mining machine and use https://mine.decisionmodels.io/troubleshooting.md.

## Related

- Use-case fine-tuning tracks (submit a model instead of serving): https://reax.co/agent/tracks.md
- Public network data (CC-BY-4.0): https://huggingface.co/datasets/decisionmodels/reax-sn92-network
- Developer docs: https://reax.dev/ · machine-readable miner descriptor: https://reax.dev/miner.json