REAX.docs Mainnet started · netuid 92
MenuRoadmap

Project

Roadmap

REAX is live on Bittensor mainnet. This page shows what is running and what comes next. Nothing here is a date commitment unless a date is named, and plans that affect miners change with at least 24 hours’ notice; adding a model to the inference list needs none.

Live on mainnet since 14 October 2026

REAX is Bittensor subnet 92 (SN92). Registration is open. Our validator is UID 3 on an RTX PRO 6000 (96 GB) and runs rounds about every 2 minutes. Release 1.2 serves four pools: s1-fast, s1-pro, s1-lite and s1-cpu. The current validator release is v1.2.0-rc11.post8.1; the current miner image tag is v1.2.0-rc11.post8.1-kit1 (miner kit 0.1.9). There is no public REAX testnet; practise on a local chain. The source is public at github.com/reaxlabs/reax-subnet.

Release 1.2 is the current mechanism

REAX is live on Bittensor mainnet as netuid 92 (since 14 October 2026). Details: Inference mining (1.2). The pages written for spec revision 0.6 describe the earlier design.

Stages

  1. Inference mining on the exact production models from Decision Models Live

    Mainnet since 14 October 2026. Miners serve one of four pools (s1-fast, s1-pro, s1-lite, s1-cpu) in a pinned runtime, and validators recompute their answers. Capacity scoring v2 has been live since 15 October: inside a pool, pay follows verified throughput, and since validator rc4 every miner's capacity is measured the same way. Several nodes per person are allowed (one hotkey and UID each), but many hotkeys on one machine share its measured capacity. The public miner board shows the live state. See Inference mining (1.2) and Verification & calibration.

  2. Peer-to-Peer tier Live · opt-in

    Customers of Decision Models can opt API keys in to network processing. Single-question s1-fast requests may then be answered by REAX miners that have a passing track record at our validator. Every network answer is checked against our own reference GPU. EU-tier requests never go to the network.

  3. Competitions Generalist track live since 21 October 2026 · engine track live since 20 October 2026

    The generalist decision-model track opened on 21 October 2026, 13:00 UTC and replaced the two use-case tracks. The REAX Engine track opened on 20 October 2026, 18:00 UTC. Both tracks are paused after their Friday 30 October slots; generalist was paused at 16:30 UTC. No next crowning is scheduled. Evaluation v3 applies when generalist is active. Engine measurements use the dedicated RTX PRO 6000 Blackwell machine. Practice data remains public; see Competition mining.

  4. Miner emission Live

    Since the first weight set at or after Monday 2 November 2026, 14:30 UTC, inference gets 30% and follows measured customer demand; it got 40% from the first weight set after the Engine slot on 30 October until then, and 50% before that. REAX Engine remains at 20% and generalist at 30%. The remaining share is not allocated to a track. Unpaid portions are burned. Both competition tracks are paused after their Friday 30 October slots; no next crowning is scheduled.

  5. Four pools and capacity-based scoring Live · release 1.2

    s1-fast (24 GB or more), s1-pro (mammoth-1, 48 GB or more; only RTX PRO 6000 Blackwell Server Edition measured), s1-lite (8-bit GGUF, NVIDIA 8 GB or more incl. Jetson Orin) and s1-cpu (8-bit GGUF on CPU). Inside a pool, weight is score times verified capacity.

  6. Several GPUs behind one hotkey Live · release 1.2.0rc8

    Works since release 1.2.0rc8: list the runtime sidecars of the same model in REAX_S1_FAST_RUNTIME_URLS or REAX_S1_PRO_RUNTIME_URLS. Since 26 October 2026, 20:08 UTC the per-hotkey cap is 2,560 decisions/s for s1-fast and 640/s for s1-pro. Since release 1.2.0rc9.post5, machines in different places can also share one hotkey (opt-in multinode, REAX_FEATURE_MULTINODE=1; see the miner guide). 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.

  7. Image-capable decision models Planned

    Decision models that read images, served in the network.

  8. Mining on everyday hardware Planned

    s1-lite (8 GB GPUs, Jetson Orin) and s1-cpu are live pools. They are not yet calibrated on real small devices; that calibration comes next, and measured speeds will be published.

  9. EU operator admission Planned

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

  10. Finetuning service & miner competition Planned · new spec version

    Planned: a finetuning service that produces custom fine-tuned decision models, with miners competing to produce better fine-tuned models. See Competition mining. The finetuning service itself is not yet specified.

Engineering backlog from spec 0.6

Work named in the earlier design (spec 0.6) that is not built yet: