By No Code MBA · Facts checked
- Provider
- AutoTrust AI
- Announcement date
- Status at last check
- Generally available
- Model family
- JEV
- Preceding model
- JEV-9B
- Capabilities
- Text
What changed
JEV-27B keeps a frozen Qwen3.8-27B text model for generation and reasoning, and adds a small LoRA (about 108.8M parameters) plus a decision head for one-pass true/false, multiple-choice (2 to 16 options) and 0 to 5 score answers, distilled from TypeSafe's closed Jev 1.13. AutoTrust reports a 84.07 mean across six decision benchmarks versus 83.85 for Jev 1.13, a 0.0009 expected calibration error against the teacher, and a median 137 ms per decision on one B200. It is larger and about 2.6 times slower than JEV-9B.
Access and availability
Open weights on Hugging Face at autotrust/JEV-27B under Apache-2.0. Both the decision path and normal generation run from one vLLM engine, with decisions served through the jev-decision LoRA module.
Official access documentationWhat you can build
Evaluate it for triage or routing by checking calibration on held-out labeled examples and the quality of the generation path.
Limitations
Calibration is verified only inside its 53 training domains and degrades on unseen task families, and about 7.4% of answers change when option order is shuffled. It inherits the teacher's weaknesses on multi-hop reasoning, arithmetic and counting, and most results are the authors' own runs.