StartLux Releases Open-Source Decision Model StartLux-Decision, Outperforming Jev on 31 of 38 Benchmarks

StartLux (Origin StarGlow) released the open-source decision model StartLux-Decision, launching five tiers at once: 0.8B, 2B, 4B, 9B, and 27B, covering the full range from lightweight local deployment to flagship specification. This is one of the few facts in this release that can be directly confirmed: the model is released, and the tier specifications are clear.

More attention-grabbing is the development pace. The team says it applied its self-built Auto Research method to decision model development and completed the research and validation of StartLux-Decision in 3 days. This claim comes from StartLux's own account, is a source_claim, and lacks independent third-party verification. It describes the team's own development process, not an externally reproducible evaluation result, so it should not be conflated with the performance numbers that follow.

Keeping these two things separate is the premise for understanding the entire release: specifications and licensing are verifiable release facts, while the development timeline and performance claims come mainly from the team's own account.

In the Decision Index 0.2.1 evaluation, the 27B version scored 63.88, above Jev 1.13's 57.91, with higher scores on 31 of the 38 benchmarks included in the index. Under the same snapshot, the 9B version also surpassed Jev 1.13 with a score of 58.63. These numbers come from self-tests run by the team using leaderboard evaluation tools, not official leaderboard results; the Decision Index comparison uses the public leaderboard snapshot from September 28, 2026.

It's necessary to understand the nature of Decision Index itself: it is an index that has been corrected against a random baseline and weighted-aggregated, not a direct accuracy rate, nor a complete agent success rate. Therefore, the gap between 63.88 and 57.91 measures comprehensive performance under that index system and cannot be directly converted into an equivalent gap in task success rate.

Domain-level results further show that the lead is not uniform. The flagship version is above Jev 1.13 in three domains—language understanding (74.5 vs. 62.0), retrieval and classification (66.8 vs. 55.4), and tools and automation (82.2 vs. 75.1)—but scores 44.3 in knowledge and reasoning, below Jev's 51.4. In other words, a lead in the composite score does not mean it is stronger in all capabilities; developers still need to go back to the types of judgments they encounter most often when making a choice.

One more constraint affects how these scores should be read: the training data used the public training sets of 14 benchmarks in Decision Index, and the corresponding test questions were filtered out. This reduces the risk of directly memorizing test questions, but it also means these benchmarks are not a completely unfamiliar evaluation environment for this model.