Physix Frontier · News Briefing Card (TMTPost · Oct 4, 2026)

StartLux Open-Sources Five-Tier Decision Model, Beats Jev on 31 of 38 Benchmarks

KEY FACTS

On September 30, Shanghai-based AI company StartLux released StartLux-Decision, an open-source decision model.
The model launched in five parameter versions: 0.8B, 2B, 4B, 9B, and 27B.
The team used an Auto Research method to complete development and validation in 3 days.
The 27B version scored 63.88 on Decision Index 0.2.1, beating Jev 1.13 on 31 of 38 benchmarks.
The 27B version averaged 91.82% accuracy across seven evaluations, higher than Jev's 88.74%.

KEY DATA

63.8827B Decision Composite Score

57.91Jev 1.13 Decision Composite Score

91.82%27B Average Accuracy Across Seven Evaluations

26ms4B Average Latency for Short Requests

PHYSIX OBSERVATION

The decision model race is shifting from parameter count to interface efficiency. StartLux uses 3-day iteration cycles and five-tier specs to lower the barrier to local deployment. Its approach of completing multiple judgments in a single request is more valuable for agent deployment than leaderboard chasing. But self-reported data diverges from official certification, so developers should select models based on their own tasks.

Source: TMTPost report