StartLux releases 27B decision model, tops Hugging Face Decision Index

StartLux, a team based in China, has released a 27B decision model built for a different job than a text-generating LLM. Rather than writing an answer, the model takes the current state and a set of candidate options and returns a probability for each one, so an Agent or another program can pick the next action.

On the Hugging Face Decision Index 0.3.2 leaderboard, updated on October 16, the 27B Instant model scored 69.12 and ranked first among 174 entries. It also placed first in all three components of the index: the public benchmark, the private test covering comparable capabilities, and the private test on new domains.

The release offers two inference paths. Instant produces probabilities in a single pass for low latency, while AutoThink uses the confidence of the initial judgment to decide whether to run additional reasoning. In the team's own published results on the public subset of Decision Index 0.3, StartLux-Decision-27B-AutoThink scored 73.37, compared with 57.96 for the hosted version of Jev, which the report describes as a relative lead of about 27%.

What has been disclosed so far stops at leaderboard placements and the team's self-reported numbers. The Instant result comes from the DI 0.3.2 leaderboard updated on October 16, while the 73.37 figure for AutoThink comes from the earlier public subset of DI 0.3, so the two scores are not drawn from the same leaderboard.