mentat-sys1-v0.1
A LoRA adapter and native decision readout on Qwen3.5-4B that answers noul, choice and score questions in one forward pass, with a probability per supplied label and a frozen scalar temperature. Self-hosted, it serves the TypeSafe /v1/systemone wire format.
A personal release by Yun Qu, with weights on Hugging Face and serving code in the mentat-asi GitHub organisation. The repo holds the adapter (rank 64, about 488 MB) and the readout, not the base weights. The temperature was fitted on a separate 1,120-row reserve set; responses add unknown-probability and abstained fields. The serving code is adapted from the Apache-2.0 Imajev project; the weights are stated to be trained independently. By the maker's own statement the training-data source and licence ledger is incomplete. The numbers are the maker's run on JevBench's public items (a third-party board), which the maker calls local measurements, not leaderboard claims; JevBench has not run it yet. No hosted API.
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 | 4B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | Yun Qu |
| Released | 2026-10-04 |
| License | apache-2.0 |
| Reported accuracy | 86.6% |
| Reported latency | 90 to 91 ms p50 / 332 to 336 ms p95 serial on one A800 80GB |
Get the weights
pip install systemonemodels
systemone pull yun-qu/mentat-sys1
The files are served from the maker's Hugging Face repository, yunqu/mentat-sys1-v0.1, and verified against the checksums recorded here.
Read more
This page was opened by System One for Yun Qu, who can claim the organisation and take it over at any time.