AutoTrust AI's JEV-35B Turns Open-Source Decisions Into 204ms API Calls
AutoTrust AI's open-weights JEV-35B turns a Qwen3.5 MoE into a System 1 decision model that picks among 256 options in 204 ms.
Takeaways
- AutoTrust released JEV-35B, an Apache-2.0 System 1 decision model fine-tuned from Qwen3.5-35B-A3B.
- Handles yes/no, 0-5 scores, and choice over up to 256 options via
/v1/decide. - Scores 59.73 on Decision Index 0.3 public suite, up from 53.64 for JEV-27B-VL.
- Median latency 204 ms on one B200; NVFP4 version fits on a 32 GB GPU.
- Strong at retrieval, classification, and tool selection; weak at reasoning benchmarks and sequential games.
- Computer-use agents reach 95% success on a 60-task browser benchmark with element text.
AutoTrust AI Lab has released JEV-35B, an Apache-2.0 open-weights model built to choose among bounded options. A request supplies a situation, a question, and as many as 256 candidate answers; the model returns its selection and a calibrated probability distribution. AutoTrust reports median single-request latency of 204 milliseconds on one NVIDIA B200.
JEV-35B fine-tunes Qwen3.5-35B-A3B, a mixture-of-experts model with 35 billion total parameters and about 3 billion active for each token. That sparse architecture limits the computation required per request. AutoTrust packages the weights with a vLLM server and a dedicated /v1/decide endpoint.
One endpoint, three decision types
Classification, routing, reranking, tool selection, and interface control share the same request shape: a typed choice over bounded options. The dedicated endpoint avoids generated JSON, schema validation, and token-by-token decoding.
| Kind | Output | Typical use |
|---|---|---|
noul |
Probabilities for false and true |
Policy checks, gating, and binary control |
score |
An integer from 0 through 5 | Quality, relevance, or risk ratings |
choice |
One of 2 to 256 supplied strings | Routing, ranking, and tool selection |
For each request, the path AutoTrust calls System 1 reads the hidden state at the final prompt token. A 264-slot decision head emits one raw score, or logit, for each slot: two binary slots, six rating slots, and 256 choice slots. The server selects the relevant subset, converts those logits into probabilities with softmax, and adjusts their sharpness using a per-kind temperature from a calibration file.
The decision-head weights are merged into rows of the language model head assigned to reserved slot tokens labeled from A through JU. This layout lets vLLM load one standard checkpoint and removes a separate LoRA adapter from deployment.