AutoTrust AI Releases JEV-27B for Self Hosted Agent Decisions

AutoTrust AI announced in a press release the release of JEV-27B, an Apache-2.0 open weights model for structured decisions in self hosted AI agent workflows. It answers yes or no, multiple choice, and zero to five rating questions in one forward pass, returning probabilities for each option.

JEV-27B adds a 108.9 million parameter decision block to a frozen Qwen3.8-27B model. AutoTrust AI says it runs on one NVIDIA B200, with median decision latency of 137 milliseconds and throughput of about 130 decisions per second.

The company reported an average score of 84.07% across six benchmark groups. It noted that comparisons with TypeSafe Jev 1.13 and published baselines were not independently validated under controlled conditions.

The model weights, decision adapter, training and serving code, vLLM support, and evaluation reports are available on Hugging Face. AutoTrust AI lists limitations involving multistep reasoning, arithmetic, dates, adversarial inputs, and its focus on English training data.