AutoTrust AI releases open JEV-27B model for self-hosted agent decisions

AutoTrust AI released JEV-27B, an Apache-2.0 open-weights model designed to handle structured decisions in AI agent workflows on one NVIDIA B200 GPU. It adds a 108.9-million-parameter decision block to a frozen Qwen3.8-27B model. That block returns probabilities for yes/no, multiple-choice and rating answers in one forward pass, while other requests use the underlying model's generation path. The release includes weights, training and serving code, and vLLM support. AutoTrust measured its own benchmark scores and latency; the company says the model struggles with arithmetic, dates and adversarial inputs and is unsuitable for high-stakes decisions.