SelfJev-4B Vision
An independent open decision model with Jev's request format: a rank-64 LoRA on Qwen3.5-4B that reads text or an image once and scores yes/no, one-choice, ordered-score and multi-select questions from forward-pass logits, returning per-option probabilities.
A LoRA adapter of about 230 MB; the Qwen3.5-4B base downloads separately. Served by the maker's selfjev package, which shares the prefix across questions and candidates; merged checkpoints also run on vLLM, but images need the maker's server. It accepts Jev-shaped requests, and the maker says code written for TypeSafe's SDK runs unchanged. Half the training targets are Jev API probabilities and half AI-checked labels. Not calibrated, per the maker. No licence is declared for the fine-tuned weights; the code is Apache-2.0. The maker's comparisons are on its own suites, which it says favour SelfJev. Siblings: the text-only selfjev-4b, merged checkpoints and GGUF builds.
What it decides
- noul — answers a yes/no question with one probability
- choice — picks one option from a set
- score — places the input on an ordered scale
- 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 | Julien Wuthrich |
| Released | 2026-09-30 |
| License | Weights: no licence declared, treated as non-commercial; code Apache-2.0 |
| Reported accuracy | 96.1% |
| Reported latency | 73 ms median, one text question, one warm request on an NVIDIA L40S |
Get the weights
pip install systemonemodels
systemone pull julien-wuthrich/selfjev
The files are served from the maker's Hugging Face repository, Jwuthrich/selfjev-4b-vision, and verified against the checksums recorded here.