JEV-27B
AutoTrust's student of TypeSafe Jev 1.13. A frozen Qwen3.8-27B plus a 108.9M-parameter decision block trained on Jev's own output distributions; one set of weights answers typed questions in one pass (System 1) or generates text with the untouched base (System 2).
Noul, Choice over 2 to 16 options, and Score on a 0 to 5 scale, served with vLLM. On a held-out set labelled mostly by Jev 1.13 itself, mean KL to Jev is about 0.017 and ECE 0.0009; the maker's run of JevBench's 231 public items gives a family-macro 88.7 against 87.2 for its own run of Jev; median 137 ms per decision on one B200. HumanEval stays at 78.0% because the base is not modified. Trained from the Apache-2.0 corpus SargeDev/jev-distill-corpus-v3, i.e. distilled from Jev's outputs; AutoTrust states no affiliation with TypeSafe. Check TypeSafe's terms before mirroring or recommending it. Siblings: JEV-9B (Qwen3.5-9B), JEV-Gemma4-26B-A4B and JEV-27B-VL.
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 calibrated probability
At a glance
| Parameters | 27B |
| Base model | qwen/qwen3.8-27b |
| Maker | AutoTrust AI Lab |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 88.7% |
| Reported latency | 137 ms median per single decision on one B200; 4.2 ms per decision in vLLM batches of 128 |
Get the weights
pip install systemonemodels
systemone pull autotrust-ai/jev-27b
The files are served from the maker's Hugging Face repository, autotrust/JEV-27B, and verified against the checksums recorded here.