StartLux-Decision
StartLux's open typed decision models from 0.8B to 27B plus a 35B-A3B MoE, reading text, JSON or images up to 256K tokens.
Overview
StartLux-Decision is a family of typed decision models: five dense sizes from 0.8B to 27B and a 35B-A3B mixture of experts. You send a state (text, JSON or images, up to 262,144 tokens at every size) and questions: pick one option, yes/no, or a rating. Every option gets a probability read from the option letters after one forward pass, so short questions take milliseconds and probabilities can serve as confidence. Requests and responses use TypeSafe's /v1/systemone format, so Jev clients work unchanged. The repository holds inference code, evaluation scripts, results and raw game logs; weights are on Hugging Face and ModelScope.
The authors report 63.88 on Decision Index 0.2.1 versus 57.91 for Jev 1.13 (the top public board entry), and that the 27B scores higher than Jev on 31 of 38 index benchmarks, while noting their runs are not on the board and their training data includes public benchmark material. Demos show the 27B ordering a product in a web store, checkmating Jev 1.13 in a Scotch Game, clearing Super Mario Bros. 1-1, beating StarCraft II's Hard AI, playing Doom and running a football team. Code is Apache-2.0; weights are CC BY-NC 4.0.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option fits (e.g. next move among eleven in Mario)? | — |
| Yes / no | Yes/no questions about text, JSON or images | — |
| Score | Rating-scale questions | — |
How it works
- Load a StartLux-Decision checkpoint
- Send state and typed questions in /v1/systemone format
- Read per-option probabilities after one forward pass
- Use probabilities as confidence for thresholds
Strengths
- Wide range of sizes including MoE
- Images and 256K context
- Drop-in for Jev clients
Caveats
- Weights are non-commercial (CC BY-NC 4.0)
- Self-reported scores not on the public board
- Training data includes public benchmark material
Getting started
Download weights from the Hugging Face collection and follow docs/inference.md.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch inference server
- Has code
- yes
- License
- Apache-2.0 (code), CC BY-NC 4.0 (weights)
Use cases and integrations
How curated lists describe it
StartLux's typed decision models in five dense sizes (0.8B to 27B) and a 35B-A3B mixture of experts, reading text, JSON, or images up to 256K tokens behind `/v1/systemone`; the authors report 63.88 against 57.91 for Jev on Decision Index 0.2.1. Code is Apache-2.0; weights are CC BY-NC 4.0.
AnotiaWang/awesome-decision-models
Frequently asked questions
How does it compare to Jev?
The authors report 63.88 versus 57.91 on Decision Index 0.2.1, from their own runs.
Can I use the weights commercially?
No; weights are CC BY-NC 4.0.
Does it work with Jev clients?
Yes, it uses the /v1/systemone request and response format.
Related projects
- jev-visual★ 317
Educational MLX project scoring typed questions about one image with Qwen3.5-0.8B on Apple Silicon, with game demos.
ModelresearchPython PlayJev: Qwen3.5-0.8B fine-tuned to play ten browser games from raw pixels, one forward pass per move with a probability per option.
ServiceresearchPythonFree hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPythonPython library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPython