lev
An open System One model from Interfaze AI. A rank-32 LoRA on Qwen3.5-4B reads yes/no, choice and score answers from the logits of one forward pass, returns calibrated probabilities and serves the /v1/systemone API.
On all 13 S1Bench subsets (Bespoke's public manifests) lev scores 68.9% macro against 76.1% for Jev in the same harness, with a mean calibration error of 0.115 (Jev 0.091). On a held-out split of its 29 training sources it reaches 80.7% (calibration error 0.061), 98.0% on banking77 and 96.8% on clinc_oos. The held-out 80.7% and its calibration error were measured before the last serving update; banking77 and clinc_oos come from after it. Engine compute is 69 ms on an H100 for a short request; S1Bench's longer states take more. Label-token readout covers choice sets of several hundred options (codes that split into two tokens are skipped), and only longer menus move to a candidate-path head, so there is no fixed option limit. English only. The adapter is Apache-2.0, but some of its training datasets carry non-commercial terms; check them before commercial use.
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
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 4B |
| Base model | qwen/qwen3.5-4b |
| Maker | Interfaze AI |
| Released | 2026-09-24 |
| License | apache-2.0 |
| Reported accuracy | 68.9% |
| Reported latency | 69 ms engine compute on an H100 (short request); 414–654 ms end to end on Modal |
Get the weights
pip install systemonemodels
systemone pull interfaze-ai/lev
The files are served from the maker's Hugging Face repository, interfaze-ai/lev, and verified against the checksums recorded here.