Open source · Sep 30, 2026
Shanghai AI company StartLux open-sources StartLux-Decision decision models in five sizes, outperforming Jev 1.13 on 31 of 38 benchmarks
Shanghai-based StartLux (原点星辉) has released StartLux-Decision, open-source decision models in 0.8B, 2B, 4B, 9B and 27B sizes that return a choice and probability instead of text, for agents in support, web, office and game settings. Raw weights and BF16, Q8_0 and Q4_K_M GGUF files support local deployment; no licence was specified.
On Decision Index 0.2.1 the 27B version scores 63.88, topping Jev 1.13 (57.91) on 31 of 38 benchmarks. On seven tasks used by Shanghai AI Lab it averages 91.82%, versus 88.74% for Jev and 90.02% for Intern-Decision-4B. The company attributes the three-day build to its Auto Research pipeline.
Original sources (Chinese)
RSI再创奇迹！StartLux推出开源决策模型StartLux-Decision，38项基准中31项高于JevJev 之后，中国团队开始深挖 AI 的「直觉层」Jev被请下王座，StartLux中国开源决策模型冲上第一