AutoTrust's Open JEV-9B Makes Fast AI Decisions at 2.5ms per Choice
JEV-9B is an open-weights 9B model combining calibrated System 1 decisions and System 2 reasoning in one engine, via Hugging Face Mean KL divergence of about 0.019 nats from closed TypeSafe Jev 1.13's output distributions on 25,376 held-out questions Trained with Blocks of Experts: 40.2M-parameter
Takeaways
- JEV-9B is an open-weights 9B model combining calibrated System 1 decisions and System 2 reasoning in one engine, via Hugging Face
- Mean KL divergence of about 0.019 nats from closed TypeSafe Jev 1.13's output distributions on 25,376 held-out questions
- Trained with Blocks of Experts: 40.2M-parameter LoRA plus 24-slot head, in about 3 B200-hours
- Median 90 ms per decision on one B200, roughly 15 times the hosted API's measured throughput
- HumanEval unchanged at 70.7% with all 164 completions byte-identical to Qwen3.5-9B base
- Now extended to vision for robot-arm control and computer-use via training corpus plus Qwen3.5 vision encoder
AutoTrust AI has released JEV-9B , an open-weights model that combines a fast decision head with the generation capabilities of Qwen3.5-9B. The decision path returns a probability distribution in one forward pass, while the frozen Qwen backbone handles ordinary text generation and step-by-step reasoning. AutoTrust trained the model to reproduce the full probability distributions of TypeSafe’s hosted Jev 1.13 model, giving developers a measurable view of how closely the open model matches its closed teacher.
A 40-million-parameter switch
Each request selects one of two execution paths in the same serving engine:
| Path | Input | Output |
|---|---|---|
| System 1 | noul for yes or no, choice for 2 to 16 options, or score from 0 to 5 | A calibrated probability distribution from one forward pass |
| System 2 | Ordinary prompts and conversations | Free-form generation using Qwen3.5-9B |
AutoTrust implements the two paths with its Blocks of Experts design. Qwen3.5-9B remains frozen as the generation block, and a detachable System 1 block supplies the decision capability. The added block contains 40.2 million trained parameters, about 0.5% of the backbone, and took roughly three hours to train on one Nvidia B200.
Keeping the blocks separate preserved the generation path in AutoTrust’s HumanEval test. The base model and JEV-9B each scored 70.7%, and all 164 completions were byte-for-byte identical. Merging the same specialist block into the backbone reduced the score to 61.6%, a loss of 9.1 percentage points.
Copying the teacher’s confidence
AutoTrust distilled System 1 from the probability distributions returned by TypeSafe Jev 1.13. The evaluation used 25,376 held-out questions spanning 53 domains, with the hosted model’s distributions serving as targets. JEV-9B recorded a mean Kullback-Leibler divergence of about 0.019 nats, where zero indicates identical distributions.
KL divergence measures how much statistical information separates two probability distributions. At 0.019 nats per sampled decision, an observer would need about 54 independent decisions to accumulate an expected one nat of evidence about which model produced them. That estimate describes imitation fidelity and does not measure correctness against human labels.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
