StartLux open-sources decision models outscoring Jev on benchmark index

Shanghai startup StartLux released StartLux-Decision, an open-source model family whose largest 27B edition scored 63.88 against 57.91 for TypeSafe AI's Jev 1.13 on Decision Index evaluations.

Why it matters

By open-sourcing weights small enough for local hardware, StartLux shifts agent routing from paid cloud endpoints directly onto client machines.

The details

• The model family spans five parameter tiers down to 0.8B, with quantized packages designed to run on personal computers.
• In direct chess matchups without search augmentation, the flagship edition won 35 out of 36 games against Jev.
• The 4B version compresses to roughly 2.71GB under Q4 quantization while maintaining a 98.3% consistency rate with unquantized weights, answering three questions in 26 milliseconds on a single H200 GPU.

The catch

The benchmark scores come entirely from StartLux's internal testing of a leaderboard snapshot, while Chinese commentators argue that standalone decision models offer little defensibility because competitors can replicate them within weeks.