StartLux-Decision v1.1 takes #1 on the official Decision Index with a non-commercial 27B

Last updated: October 16, 2026

publicado (OK do Fabricio 17/10 BRT).

StartLux Labs released StartLux-Decision v1.1 on 10-11 October and posted a technical report on 12 October. The family has two variants: Instant (0.8B to 27B, one forward pass) and AutoThink (4B, 9B, 27B), which reads the answer in one pass and writes a short thought only when its top probability is under a 0.6 gate. Weights are CC BY-NC-4.0; the serving package, with a /v1/systemone endpoint and GGUF and MLX builds, is Apache-2.0.

Official result. On the Decision Index 0.3.2 (data of 16 Oct 19:00 UTC) the 27B scores Full 69.12, first and not tied, ahead of vLLM-SR d3 (66.26), linne-oak (65.61) and Perplexity Decider v1.2 (65.55); Jev 1.13 has 60.11. On JevBench open weights the 27B Instant has Intelligence 73.2 and calibration 91.9 with a measured p50 of 55 ms, though its cost puts it outside the Jev-class cap.

Caveats. The cards say the authors picked training recipes by the Decision Index public subset, so the public score (72.83 own run, 72.88 on the board) is a model-selection number; the held-out parts of the Full score are what to read. The DecisionPivot benchmark the authors introduce (AutoThink 86.9 pair accuracy) is their own and unverified. Commercial use needs a license from StartLux Labs.

The StartLux-Decision page was re-scored from 5.1 / 7.0 / 5 to 5.3 / 8.3 / 11 on this evidence.