StartLux open-sources decision model that tops Jev on public benchmarks

StartLux, a Shanghai AI company, has released StartLux-Decision, a fully open-source decision model family with 0.8B, 2B, 4B, 9B, and 27B variants plus quantized files for local deployment. According to the team’s self-test using public evaluation tools and a Sept. 28, 2026 snapshot of the Decision Index leaderboard, the 27B model scored 63.88 on Decision Index 0.2.1, ahead of Jev 1.13 at 57.91, and outperformed Jev on 31 of 38 benchmarks. In a separate seven-test suite from the Shanghai AI Lab Intern-Decision team, the model posted an average accuracy of 91.82%. StartLux also shared a head-to-head chess comparison in which both models saw the same board state, used no extra search, and chose each move directly. The company said StartLux-Decision-27B won 35 of 36 games and performed better on average loss, best-move hit rate, and blunder count. The release is being framed not just as a benchmark result, but as part of StartLux’s broader local AI stack, where a dedicated decision model handles high-frequency, low-latency choices inside agent systems while larger models take on more complex reasoning.