AutoTrust AI Debuts JEV-27B for Self-Hosted Decision Making

The model uses a technique AutoTrust AI calls its "Blocks of Experts" architecture. A lightweight, 108.9-million-parameter decision block—comprising just 0.4% of the total model—is trained to handle System 1 tasks like yes/no queries and probability-calibrated ratings. Because the core reasoning path remains untouched, the model preserves its original generative capabilities. Training for the decision block required approximately 9.2 hours on a single NVIDIA B200 GPU, ensuring the system remains accessible for private, self-hosted infrastructure.

Performance metrics show the model achieving an 84.07% average across six benchmark groups. While internal testing suggests it performs competitively against the hosted TypeSafe Jev 1.13 API, the company notes these results are internal comparative evidence rather than independent validation. According to CEO Daniel Tang, the shift toward localized decision-making fundamentally alters the cost and risk profile for companies that require data privacy for agentic workflows. The model is currently available under an Apache-2.0 license, with support for vLLM and full serving code included in the release.