REAX 1.1: a 27B decision model fine-tuned from pplx-decider
Apache-2.0 weights built on REAX 1, itself a fine-tune of pplx-decider-v1.2. The card reports 71.27 on its own SN92 scoring.
REAX published REAX 1.1 on Hugging Face on 21 October 2026, by repository creation date. The card describes a 27B decision model that reads a state and several questions in one forward pass and returns calibrated probabilities without generating text. The maker says it is not affiliated with Perplexity AI or the Qwen team.
It is a fine-tune of REAX 1, which is a fine-tune of pplx-decider-v1.2 from Perplexity, derived from Qwen3.8-27B. REAX 1.1 combines REAX 1 with three LoRA fine-tunes. About 26B parameters, BF16, Apache-2.0. The training dataset is listed as CC-BY-4.0.
On what the card calls SN92 REAX scoring over six pool draws, pplx-decider-v1.2 scores 65.34, REAX 1 scores 67.65 and REAX 1.1 scores 71.27. The paired gain over REAX 1 is +3.62 with a standard error of 0.48. Internal safety classification is 85.98%.
The card lists its own limits. It was trained on requests up to about 12k tokens. Part of the safety result is in-distribution. It is not ranked on JevBench, and the Hugging Face and AA Decision Index submissions still name REAX 1. High-stakes uses need human review.
The release is a day old at the time of writing. The card is thorough about what it has not shown.
Sources
Editor's first pass (AI-assisted)
- Real 3.5: Public weights; it cites a dataset under CC-BY-4.0.
- Useful 3: General decisions with a safety component; scored on a validator the card admits is not on JevBench.
- Notable 2.5: Shows the stacking of open fine-tunes.
- Novel 2.5: A fine-tune of a fine-tune.
Comments