RSI-Jev v6.0-VL 4B
Jev-style decision model fine-tuned from Qwen3.5-4B-Base by a self-improving loop of AI agents. Answers noul, choice and score questions about text and up to four images in one pass, with a calibrated probability per option; effort picks an exit at layer 16, 20 or 32.
A personal project of Shanghua Gao. An agent loop trained, evaluated and documented each release: seven releases in twelve days, with hundreds of experiments logged, failures included. No generated text; effort low, medium, high or auto picks the exit depth, and image questions always use all 32 layers. Ships a Jev-compatible HTTP server and a pip package. Training data includes benchmark train splits; the maker found about 1,000 Decision Index test items in his corpora and re-scored without them, and his own run of that third-party board gives 46.24. His internal 15-benchmark suite (0.770) overlaps training data, so the held-out number below is the cleaner one. The weights are Apache-2.0, but the vision releases were trained on several non-commercial or research-only image datasets. Siblings: v5.0-VL 3B, v4.0-VL 2B and earlier text-only releases.
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 | 4.69B |
| Base model | Qwen/Qwen3.5-4B-Base |
| Maker | Shanghua Gao |
| Released | 2026-10-06 |
| License | apache-2.0 |
| Reported accuracy | 69.8% |
Get the weights
pip install systemonemodels
systemone pull shanghua-gao/rsi-jev
The files are served from the maker's Hugging Face repository, shgao/rsi-jev-v6.0-vl-4b, and verified against the checksums recorded here.