AutoTrust’s JEV-27B-VL lands on inference.sh

AutoTrust said its Hugging Face launch article for JEV-27B-VL is now live on inference.sh’s app listing and that the model is free to run there for a limited time. In an X post, the company said the listing came through a partnership with inference.sh and that the two companies plan co-marketing in the coming weeks.

The inference.sh listing describes JEV-27B-VL as an AutoTrust open decision model with vision. It lets users send text and up to eight images, then ask typed choice, score or yes/no questions. Rather than generating text, the model returns calibrated probabilities for every answer option.

AutoTrust originally published the model card on Hugging Face on September 30, 2026, describing JEV-27B-VL as "JEV-27B with vision." The company said the open-weight, Apache-2.0 licensed model is a multimodal decision system that accepts text and images and produces a calibrated probability for each option in a single forward pass.

What does JEV-27B-VL do?

JEV-27B-VL is designed to make structured decisions from text and images instead of writing long responses. AutoTrust said the model has two modes: System 1 for fast calibrated probabilities, and System 2 for slower step-by-step reasoning while looking at images.

Inference.sh’s own announcement said the model can make instant decisions on multiple questions in parallel and evaluate images, while also noting that access would be free until the end of October, subject to capacity limits. The platform also said JEV-27B-VL was #2 on the Hugging Face JEV index for vision decision models.

AutoTrust said the model’s decision head never saw an image during training. In the company’s framing, that architecture lets the system reason over vision inputs without turning the workflow into text generation.

What results did AutoTrust report for the model?

AutoTrust said JEV-27B-VL’s benchmark results were strongest in its own launch materials. The company said System 1 scored 89.9 on RewardBench, ahead of published judge results for Gemini 1.5 Pro, GPT-4o and Claude 3.5 Sonnet.

The company also said a confidence filter raised trivia-answer accuracy from 71% to 96% on the half of questions it trusted most. For retrieval and recommendation, AutoTrust said its zero-shot news recommendation result on MIND reached an AUC of 0.642 and exceeded listed zero-shot baselines and two LightGBM rankers trained on the dataset.

AutoTrust said the model reached nDCG@10 of 0.858 on TREC-COVID, above bge-reranker-v2-m3 and BM25. In a MicroLens test with 200 users and 20 candidate videos per decision, the company said JEV-27B-VL using cover images only achieved an AUC of 0.727 and a 59% top-5 hit rate, while collaborative filtering scored AUC 0.728 and a 49% top-5 hit rate.

How is AutoTrust pitching the model’s use cases?

AutoTrust is presenting JEV-27B-VL as a decision model for several structured workflows. The company said the model matched collaborative filtering on one image recommendation test while using no interaction data, and added that cover images outperformed titles alone in that benchmark.

On text-only decisions, AutoTrust said JEV-27B-VL gave the same answers as JEV-27B on all 24 reference decisions it checked. The company also said the model can be served with vLLM on one GPU with 80 GB or more, which it framed as making deployment possible on its own infrastructure.

AutoTrust said its demo package includes code and an interactive web app covering image recommendation, news recommendation, judging and hallucination guard use cases. It also described JEV-27B-VL as open weights under an Apache-2.0 license.