REAX 1.3 enters Decision Index 0.4.1 at #1 with a Full score of 73.98
REAX 1.3 (decisionmodels.io, Bittensor subnet 92; Apache-2.0, repo created 1 November 2026) is a decision model in two parts that run side by side on one RTX PRO 6000 (96 GB). Part G is a miner fine-tune on the Gemma line blended with KnowLine-31B-Gen1; part Q is a miner fine-tune on the Qwen3.5-27B line blended with d3-ultra. Each part returns a probability for every option, and the engine averages them 50/50. Weights are about 115 GB in FP8; serving uses the MIT REAX Engine with POST /v1/systemone.
Board row. In the Decision Index 0.4.1 data generated on 6 November 2026 (09:05 UTC, 274 entries), REAX 1.3 has a Full score of 73.98, #1, ahead of glide-v1.2.1-loop (73.37). Public index 73.71; ECE 0.018; median latency 246 ms with p95 2.7 s. The board lists it as an inference technique of about 57B served parameters, not a single checkpoint. The card’s own figures (public 73.71, median 207 ms single-request) are the vendor’s.
Caveats. It is an ensemble of other people’s fine-tunes, credited in the repo, and needs a 96 GB GPU, 400 GB of disk and minutes of warm-up. REAX’s maintainer also runs JevBench and keeps REAX models off that board. Hugging Face showed 0 downloads and 5 likes on 7 November. No catalog page yet: REAX 1 and 1.3 stay as news until the maturity and adoption scores can be worked out from evidence. No score change.
12 Nov 2026 file. data/index.json generated 2026-11-12T12:34:44Z (300 entries) has REAX 1.3 at Full 72.64, strict third, behind karpo-decision-31B (73.86) and GLiDE 27B (73.37). The public index in that file is still 73.71. Median 245.9 ms, ECE 0.0179. The 73.98 above is the 6 Nov file, not this one. Still no catalog page.
