StartLux-Decision-4B
StartLux's decision family in five sizes, 0.8B to 27B. All questions in a request are answered in one forward pass, with a probability for every option, through a TypeSafe /v1/systemone-compatible server that records CUDA graphs for short requests.
The card does not name a base model; the configuration is a Qwen3.5-architecture checkpoint. The 4B answers 204 of JevBench's 231 public items (as bundled by Intern-Decision) and takes 26.0 ms for a three-question request, 14.7 ms for one yes/no question, end to end on one H200 in BF16. StartLux also quotes a Decision Index 0.2.1 score of 52.75 from its own run; the model was not on the public board as of 28 September. The 27B answers 208 of 231. Weights are CC BY-NC 4.0 (commercial use needs a licence from StartLux, so do not mirror); the inference code is Apache-2.0. GGUF builds exist for every size.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
At a glance
| Parameters | 4.7B |
| Maker | StartLux |
| Released | 2026-09-29 |
| License | cc-by-nc-4.0 |
| Reported accuracy | 88.3% |
| Reported latency | 26.0 ms per three-question request, 14.7 ms per yes/no question, end to end over HTTP on one H200 (BF16) |
Get the weights
pip install systemonemodels
systemone pull startlux/startlux-decision
The files are served from the maker's Hugging Face repository, startlux-models/StartLux-Decision-4B, and verified against the checksums recorded here.