Skip to main content

AutoTrust AI Launches JEV-27B Decision Model for Self-Hosted Agents

Image
AutoTrust AI Launches JEV-27B Decision Model for Self-Hosted Agents

SINGAPORE – September 28, 2026 -- AutoTrust AI released JEV-27B, an open-weights decision model that trained in roughly 9.2 hours on a single NVIDIA B200 GPU, targeting enterprises that need fast, calibrated AI decisions without sending data to third-party APIs.

Decision block adds 108.9 million parameters to a frozen backbone

JEV-27B pairs a 108.9-million-parameter decision block, about 0.4% of the total model, with a frozen Qwen3.8-27B backbone from Alibaba. The model answers yes/no, multiple-choice and 0-5 rating questions in a single forward pass, returning calibrated probabilities for each option, while preserving the base model's full generation and reasoning path. With the decision block disabled, all 164 HumanEval completions matched the base model byte-for-byte, and the model scored 78.0% on the coding benchmark.

Benchmarks show 84.07% mean accuracy across six test groups

AutoTrust AI reported scores of 88.70% on JevBench, 83.75% on Kev, 73.89% on OpenJev text, 92.91% on Nimble, 77.46% on VitaminC and 87.71% on MASSIVE-en, for an equal-weight mean of 84.07%. In AutoTrust's own comparison, the hosted TypeSafe Jev 1.13 API scored a mean of 83.85% on the same tests, ahead on two groups and behind on four; the company said the figures should not be read as independent third-party validation. On an independent test scored against human labels, decision-models-under-pressure, JEV-27B reached 96% of Jev 1.13's accuracy across 16 answer options, without exceeding it. Distillation fidelity testing on 25,376 held-out questions produced a mean KL divergence of 0.017 against Jev 1.13's probability distributions.

Local inference cuts latency to 137 milliseconds per decision

Running on a single B200, JEV-27B posted a median decision latency of 137 milliseconds and throughput of about 130 decisions per second. AutoTrust AI's model card cites third-party measurements of 238 to 301 milliseconds and 23 decisions per second for Jev's hosted API, though the company noted the hosted figures include network time and differing hardware and concurrency conditions, making the comparison not str

Published by
fairsonline_team