StartLux-Decision Update: 27B AutoThink Hits 73.37, Claims Decisive SOTA Lead
新智元 · wechat · 2026-10-16
StartLux released an updated decision model family claiming top scores across all sizes on DecisionIndex 0.3. The 27B AutoThink variant scores 73.37, far ahead of Jev (hosted, 57.96), Cloudflare Clef (61.71) and Perplexity Decider (62.25).
Two branches: Instant models use bidirectional attention in a single forward pass (up to +2.4 points in 4 of 5 sizes), while AutoThink adds confidence-guided adaptive reasoning with up to 4096 reasoning tokens. The original RLCD² method (Reinforcement Learning for Calibrated Deliberative Decisions) rewards improvements in the decision distribution plus a soft length penalty, raising 9B AutoThink's proper termination rate from 45.9% to 93.9% and cutting average reasoning length 65%.
The team also open-sourced the DecisionPivot benchmark (13 task families, 7,800 paired-control questions); 27B AutoThink reaches 86.9% pairwise accuracy. Weights and inference code are public (Apache-2.0 code, CC-BY-NC 4.0 weights).
More from Models
- Open models now above Sol-level, outpacing expectations as community awaits V4.1 Pro — yacineMTB · 2026-10-18
- Dev's model vibes: Opus 5.5 writes best, Sol 6.1 best value, Sonnet skippable — ricklamers · 2026-10-18
- Codex took 3 days, Opus 5.5 nailed it in an afternoon — banned user's video protest — dotey · 2026-10-18
- Yacine MTB: Users will remember how Anthropic treated them once intelligence proliferates — yacineMTB · 2026-10-18
- Mistral's Timestral forecasting models start at 5M params and run on plain CPUs — lmoroney · 2026-10-18
- AutoTrust's Trending HF Models Powered by Thousands of Farmed Likes — mrfakename0 · 2026-10-18