AutoTrust AI's JEV-35B Replaces Entire Routing Pipelines With One 204ms Decision Model

AutoTrust AI Lab has released JEV-35B, an open-weights model that returns calibrated probabilities over fixed options. It uses Qwen3.5-35B-A3B, a mixture-of-experts base with roughly 3 billion active parameters per token. Its Hugging Face repository lists more than 1.6 million downloads.

AutoTrust positions the model as a replacement for pipelines assembled from retrievers, classifiers, routers, and rerankers. Each request supplies a state, a question, and as many as 256 options. The release reports an overall median latency of 204 milliseconds on one Nvidia B200.

A fixed output surface

The POST /v1/decide endpoint supports three decision kinds:

A 264-slot decision head produces one raw score, or logit, per slot. The server applies softmax only to the slots used by the requested decision kind, with separate temperature calibration for each kind. The head is merged into lm_head rows assigned to slot tokens, allowing vLLM to load one checkpoint without LoRA adapters.

Send a decision request

curl http://localhost:8000/v1/decide \
  -H 'Content-Type: application/json' \
  -d '{
    "kind": "choice",
    "state": "Customer message: my card was charged twice for the same order.",
    "question": "Which team should handle this ticket?",
    "options": ["billing", "shipping", "technical support", "account security"]
  }'

The response returns the options, their probabilities, the highest-probability choice, and token usage. JEV-35B reads the answer from the hidden state at the final prompt token, so the model never generates the response syntax. This design removes malformed model-generated JSON and conversational preambles from the decision path.