decisionmodels/REAX-1.3

REAX 1.3

REAX is a decentralised lab on Bittensor subnet 92: miners train decision models in an open competition, and our releases build on the best of them.

REAX 1.3 is a decision model with two parts that run in parallel on a single GPU (reference: one RTX PRO 6000). Each part gives a probability for every option of every request, and the model combines these probabilities into an answer. It returns probabilities over typed decision options through a Jev-compatible interface.

The two parts are:

• Part G: miner model uid 18 (CharlieBT/gx31-decider, a fine-tune trained by a REAX miner; Gemma line, derived from tattrongvu/Jeffan-31b-1710), blended with PelaAI/KnowLine-31B-Gen1 (Gemma-4-31B).
• Part Q: miner model uid 83 (gurujustin/sn92-gen-p5, a fine-tune trained by a REAX miner; Qwen3.5-27B lineage via baselabs/mammoth-1), blended with vllm-sr/d3-ultra (27B, Qwen lineage).

Credits, licences and notices for all upstream contributors, including the competition miners, are in ATTRIBUTION.md, NOTICE and licenses/.

Run with the REAX Engine

The inference code is the reax13 mode of the open-source REAX Engine (MIT License), directory reax13/ at commit ff8f595f149abdd50ea914455eadc74cf97a277f. The engine/ folder of this repository holds the same files, byte for byte. Every request goes to both parts; their per-option probabilities are averaged 50/50 and every supplied option is returned.

You need: Linux, one NVIDIA RTX PRO 6000 Blackwell 96 GB (driver with CUDA 13 / SM120), Docker with the NVIDIA Container Toolkit, at least 400 GB free disk and 64 GiB RAM (128 GiB recommended while the FP8 weights load).

# 1. Engine (REAX Engine, reax13 mode) and the Decision Index kit it builds on
git clone https://github.com/reaxlabs/reax-engine.git
git -C reax-engine checkout ff8f595f149abdd50ea914455eadc74cf97a277f
cd reax-engine/reax13
docker build -t reax13-base -f Dockerfile .
docker build -t reax13-runtime -f Dockerfile.conv .
git clone https://github.com/apolinario/decision-index.git decision-index
git -C decision-index checkout 25c6ec7b6b5f44a216c63c3a700de3a9c6d6699d

# 2. Weights: both parts from this revision (about 115 GB)
python3 -m venv .hf-cli
source .hf-cli/bin/activate
pip install -U "huggingface_hub[cli]"
hf download decisionmodels/REAX-1.3 --include 'part-g/*' --include 'part-q/*' --local-dir model-repo

# 3. Serve (local files only): GET /healthz, POST /v1/systemone on 127.0.0.1:8080
docker run --rm --name reax13-server -p 127.0.0.1:8080:8080 \
  --gpus device=0 --read-only --user "$(id -u):$(id -g)" \
  --cap-drop ALL --security-opt no-new-privileges --shm-size 8g --tmpfs /tmp:rw,exec,size=8g \
  -v "$PWD/model-repo:/model:ro" -v "$PWD/decision-index/decision_index:/kit/decision_index:ro" \
  --entrypoint env reax13-runtime -i \
  PATH=/usr/local/bin:/usr/bin:/bin HOME=/tmp PYTHONPATH=/engine:/kit \
  DUO_ISOLATED=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 HF_HUB_DISABLE_TELEMETRY=1 HF_HOME=/tmp/hf \
  KERNELS_CACHE=/opt/kernels LOCAL_KERNELS=kernels-community/finegrained-fp8=/opt/fp8-kernel \
  D3_STATE_ONCE=0 OMP_NUM_THREADS=4 OPENBLAS_NUM_THREADS=4 \
  PYTORCH_ALLOC_CONF=expandable_segments:True TRITON_CACHE_DIR=/tmp/triton \
  python /engine/serve.py --host 0.0.0.0 --port 8080 --model /model --config /engine/reax13.json

# 4. Ask a question (first answer after loading and warmup, a few minutes)
curl -s 127.0.0.1:8080/v1/systemone -H 'content-type: application/json' -d '{"state":"A parcel arrived late.","questions":{"route":{"type":"choice","instructions":"Which team should handle the complaint?","criteria":{"delivery":"Delivery support","billing":"Billing support"}}}}'

To score with the Decision Index kit instead of HTTP, use the python -m decision_index run --engine reax13_engine:Reax13Engine ... command in engine/README.md. Requests above 32,000 input tokens (size proxy) are refused as unsupported, never truncated; long requests can take minutes, so allow a client timeout of at least 1,800 s.

Hardware: both parts are loaded in FP8 (block-128) on that one GPU; peak reserved memory about 88.5 GiB. "Jev-compatible" describes the request/response interface only; it does not imply endorsement or identical outputs.

Measurements (our own, internal; not official leaderboard scores)

Measure	Result	Basis
Public Decision Index 0.3, full public suite, native index	73.71	Run reax13-public-0.3-20261103 (results in decisionmodels/REAX-1.3-decision-index), all 140,178 scoreable requests answered, official scorer, FP8 engine. Counting the one request above the released engine's 32,000-token cap as refused gives 73.70
Latency median / mean / p80	207 / 655 / 741 ms	Released engine, uniform 1,000-call sample of the public suite, one RTX PRO 6000, one request at a time. The latency in the run's scores.json comes from the parallel run

Long multi-question requests take far longer than the median; some take several minutes, so allow a client timeout of 1,800 s.

Comparison on public items not used in our recipe selection

Own measurement, not official scores, and not comparable with the 73.71 above. Native index with the published Decision Index 0.3 scorer on the 3,478 of 3,513 drawn public Hugging Face Decision Index items that all eight models answered (two fixed-seed draws of 1,000 and 2,513 items, 41 of 43 benchmarks). The second draw was added after a partial result on it (REAX ahead of KnowLine and d3-ultra) was known. Default settings; 95% CIs reflect item sampling only.

Model	Index	95% CI	REAX 1.3 minus this model (95% CI)	How it ran
REAX 1.3	69.5	[67.4, 71.4]		Run by us, FP8
d3-ultra	68.8	[66.5, 70.8]	+0.7 [-0.8, +2.2], within noise	Run by us
KnowLine-31B-Gen1	67.4	[65.1, 69.3]	+2.1 [+0.6, +3.8]	Run by us
d1-ultra	64.7	[62.0, 66.6]	+4.8 [+2.9, +7.1]	Vendor API
Clef	60.1	[57.2, 62.2]	+9.4 [+7.1, +12.1]	Author's server software on our rented GPU
Jev 1.14	59.5	[56.9, 61.4]	+10.0 [+8.1, +12.1]	Vendor API
Microsoft-Decision-1	55.9	[53.1, 57.9]	+13.6 [+11.5, +16.1]	Vendor API, our Foundry deployment
OpenAI Decisions (gpt-6-luna)	45.1	[42.7, 47.4]	+24.4 [+22.1, +26.6]	Vendor API

This is a manual selection, not a ranking; other models are not shown. The public Hugging Face Decision Index ranks by its Full score, which includes private test sets we cannot measure, so these numbers do not say where REAX 1.3 would rank.

Limitations and licences

Accuracy and calibration vary by task and domain. Two model parts need more GPU memory and compute than either part alone.

The Apache-2.0 root licence covers our blended work, subject to the preserved upstream licences and notices (KnowLine weights Apache-2.0, KnowLine serving code MIT, Gemma 4 Apache-2.0, gx31-decider Apache-2.0, d3-ultra Apache-2.0, sn92-gen-p5 MIT). See licenses/ and ATTRIBUTION.md; model metadata does not replace those component terms.