The Pulse
AutoTrust Releases JEV-27B, a Self-Hosted Decision Model
AutoTrust AI released JEV-27B, an Apache-2.0 model that combines fast, structured decisions with the reasoning capabilities of a Qwen3.8-27B backbone. The company reports an 84.07% average across six benchmarks, while cautioning that the co
AI.info Team · Written with AI · fact-checked ·
“JEV-27B shows that this capability can run on one GPU inside a customer's own infrastructure, next to a reasoning model.”
Daniel Tang, chief executive and co-founder, AutoTrust AI
AutoTrust AI released JEV-27B on September 28, presenting it as an open-weights model for agent workflows that need frequent, structured decisions without handing each request to a hosted service. The model answers yes-or-no questions, selects among multiple options or assigns a rating from zero to five, returning probabilities for the available answers. Its Apache-2.0 license permits self-hosting, and the release includes model weights, a decision adapter, training and serving code, and evaluation reports.
One model, two kinds of response
JEV-27B pairs a frozen Qwen3.8-27B backbone with a detachable decision block containing 108.9 million trained parameters—about 0.4% of the model. AutoTrust calls the quick, typed decision path “System 1” and the ordinary text-generation and reasoning path “System 2.” A router directs requests to either capability while both use the same set of weights.
The distinction matters for teams that want a model to classify a ticket, flag a potentially fraudulent message or assess a code change, then use a language model for a more involved response. AutoTrust says the decision block took about 9.2 hours to train on one NVIDIA B200. With that block switched off, the company says, all 164 HumanEval coding-test completions matched the base model byte for byte.
AutoTrust’s 84.07% benchmark average
Across six public decision benchmark groups, AutoTrust reports an equal-weight average of 84.07% for JEV-27B. It ran TypeSafe Jev 1.13, a closed hosted model, on the same groups and reports an 83.85% average for Jev; JEV-27B scored higher on four groups and lower on two. The individual results show why the mean needs context: Jev led on Kev and VitaminC, while JEV-27B led on JevBench, OpenJev text, Nimble and MASSIVE-en.
Those head-to-head figures come from AutoTrust’s own runs, not an independent evaluation. The company also cites a separate benchmark using human-labeled data, where JEV-27B reached 96% of Jev’s accuracy on questions with 16 answer options. That result does not show parity on every task, and AutoTrust says the model reproduces some of Jev’s weaknesses, including problems with multi-step reasoning, arithmetic and adversarial inputs.
Local speed comes with a hardware trade-off
AutoTrust reports a median of 137 milliseconds per decision and throughput of about 130 decisions per second on one B200. It compares those numbers with third-party measurements of Jev’s hosted API, but the company notes the tests are not controlled equivalents: its local timings omit network delays, while the hosted measurements include them and depend on hardware, serving conditions and concurrency.
The model can also run on a single 80-GB H100 at full bf16 precision, according to AutoTrust, though with lower capacity than on a B300. The decision model’s size and self-hosting option may suit organizations that need to keep requests on their own infrastructure, but the 27-billion-parameter backbone still requires substantial hardware. AutoTrust advises confidence-based routing and says JEV-27B is not intended for high-stakes decisions; its strongest benchmark claims remain company-run comparisons rather than third-party leaderboard results.