StartLux-Decision
Fully open-source (Apache-2.0) family of typed decision models from Chinese startup StartLux in five dense sizes 0.8B–27B plus a 35B-A3B MoE, returning per-option probabilities instead of generated text for agent judgment tasks.
Why it matters
Self-reported 63.88 on Decision Index 0.2.1, beating Jev 1.13 (57.91) on 31 of 38 benchmarks; repo added image input + 256K context on Oct 3, and the README documents real demos (web shopping, chess vs Jev, Mario, StarCraft II, Doom) running on one H200.
What you could build with it
An indie developer building a browser-automation agent could replace each step's LLM call with StartLux-Decision-4B's millisecond judgments on the next UI action, cutting per-step latency and inference cost for high-volume web tasks.
Does it hold up?
Usable now: Apache-2.0 weights, a TypeSafe-compatible /v1/systemone API, GGUF local builds, and six documented demos (chess, Mario, StarCraft II, Doom, web shopping) with raw logs. The 63.88 Decision Index score is self-reported by the vendor, not yet on the public board. Caveat: the vendor discloses its training used public train splits of 14 of the 38 benchmarks.
Built with StartLux-Decision
- StartLux's open-source decision model outperforms Jev in benchmark testsKuCoin News · In-depth analysis of the Decision Index results, the local-agent system architecture, and the Auto Research pipeline the team says built the model direction in ~3 days.
- Jev Dethroned: StartLux China Open-Source Decision-Making Model Surges to Top Spot36Kr Europe · Chinese tech-media writeup of the release covering the 63.88 vs 57.91 benchmark comparison against Jev 1.13, the chess and shopping demos, and StartLux's local-intelligence strategy.
- StartLux-Decision model weights collectionHugging Face · Official weights for all six model sizes plus GGUF quantizations; several quantized variants already show thousands of downloads, indicating real early adoption.
Learn more
First spotted on github: source.
More AI models
ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69Mistral Forge: enterprise platform for training and continuously improving proprietary modelsMistral AI launched Forge (Sep 28) — a full-lifecycle model training platform (pre-training, SFT, DPO/ODPO, RL…models · JEV 0.69kev: open, trainable Jev-like family of small decision models on Qwen3.5/3.8A family of small decision models (0.8B to 27B) built on Qwen3.5/3.8 that reproduces Jev's prefill-only typed-decision…models · JEV 0.68GPT-Synopsys: OpenAI and Synopsys multi-year deal to build an AI model that operates EDA toolsOpenAI and Synopsys signed a multi-year agreement on Sep 30 to jointly develop GPT-Synopsys, a specialized model that…models · JEV 0.66
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.