AutoTrust JEV-27B-VL Beats GPT-4o in Image Scoring

AutoTrust has introduced JEV-27B-VL, a new vision-language model designed to make rapid, structured decisions from visual and textual inputs. Built on the Qwen3.8-27B backbone under an Apache-2.0 license, the model integrates a LoRA adapter and a specialized decision head. This architecture allows the model to bypass slow, token-by-token label generation. Instead, its System 1 interface returns calibrated probabilities for yes-or-no questions, ratings from 0 to 5, or selections among 2 to 256 choices in a single forward pass.

The model has demonstrated strong performance across several benchmarks, surpassing prominent proprietary models. JEV-27B-VL achieved a 78.3% score on VL-RewardBench and 73.2% on Plan-RewardBench, outperforming GPT-5 and Gemini-3-Flash in agent judging tasks. In practical control loops, the model achieved a 75% success rate in MuJoCo robot pick-and-place simulations and a 95% success rate across 60 multi-step browser automation tasks. Additionally, in a zero-shot video recommendation test using only cover images, it nearly matched a collaborative filtering system trained on 59,045 users, scoring an AUC of 0.727 compared to the traditional system's 0.728.

For machine learning practitioners, JEV-27B-VL offers a highly efficient alternative for classification, ranking, routing, and evaluation tasks. The model supports context windows up to 256K tokens and is deployed via a POST /v1/decide endpoint built on top of vLLM. Because the decision head outputs calibrated probabilities, developers can establish reliable confidence thresholds. This allows them to route routine, low-latency decisions through the fast System 1 process, while reserving more complex, open-ended queries for System 2 reasoning. The model's underlying text decision capabilities have already proved popular, with its Hugging Face page recording over one million downloads.