Messier One
A Qwen3.5-4B fine-tune that answers choice (up to 255 options), score and yes/no questions with a probability per option from one forward pass, read from label tokens in the language-model head, behind a TypeSafe-compatible /v1/systemone server. Non-commercial weights.
A rank-16 LoRA merged into Qwen3.5-4B. The trained readout sits in the head rows of 229 added option-label tokens, with one fitted temperature per question type; no overall calibration figure is published. On JevBench's 231 public items, in the maker's own run with JevBench's runner (not a leaderboard result), it answers 197. On its own many-option test, with reference answers from larger LLMs, it scores 97.6% at 2 to 26 options and 89.4% at 101 to 255. Option order can shift answers; arithmetic and date computation are unreliable. The weights are CC BY-NC 4.0: commercial use needs a licence from the maker, so they are not mirrored or packaged here. The serving code is Apache-2.0.
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 probability
At a glance
| Parameters | 5.3B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | agentmessier |
| Released | 2026-10-06 |
| License | cc-by-nc-4.0 |
| Reported accuracy | 85.3% |
| Reported latency | p50 46 ms easy, 51 ms standard, 95 ms hard per decision on one RTX 5090 (vLLM 0.30.0, BF16) |
Get the weights
pip install systemonemodels
systemone pull agentmessier/messier-one
The files are served from the maker's Hugging Face repository, agentmessier/messier-one, and verified against the checksums recorded here.
Read more
This page was opened by System One for agentmessier, who can claim the organisation and take it over at any time.