REAX 1: an open 27B decision model fine-tuned from Perplexity Decider v1.2, kept off JevBench

Source · Model card: pplx-decider v1.2 (27B + Flash 4B)

publicado (OK do Fabricio 21/10 BRT).

On 21 October 2026, REAX, the decision-model network behind decisionmodels.io (Bittensor subnet 92), released REAX 1. It is a 27B decision model under Apache-2.0, fine-tuned from Perplexity Decider v1.2, which is itself built on Qwen3.8-27B. It keeps the base’s joint protocol: one request carries a state and several choice, yes/no or score questions, and returns probabilities with no generated text. The card describes training as four LoRA fine-tunes (hard-mined decision questions, generated items, public SN92 specialist rows, Aegis 2.0 safety rows), merged into the backbone as a “model soup”. It is also the base model for the SN92 generalist miner competition.

Numbers are the vendor’s own. UnverifiedClaim: The card reports an internal composite of 81.39 against 79.76 for the base (+1.63), with AA-DI-style items at 64.82 against 61.96. The launch post says “+2.4 points over its base, ahead in all 6 test draws and all 10 areas”. The post and the card give different margins. Neither is a leaderboard result, and REAX 1 is not on Decision Index 0.4 or Artificial Analysis yet.

Conflict of interest, disclosed. Florian S, who runs JevBench and is behind REAX, wrote that REAX models “might be the new state-of-the-art”. He added that, “to ensure JevBench stays independent”, they will not be listed there and will go to the Hugging Face Decision Index and Artificial Analysis instead. We will score REAX 1 once one of those boards has a row for it.