Autotrust's JEV-27B-VL, listed as Autotrust-Decider-Pro-Vision-27B, is a dual-system vision decision model built on Qwen3.8-27B. System 1 uses a text-trained adapter and decision head atop an unchanged vision encoder to return calibrated probabilities over typed choices—yes/no, 2–256 options, or 0–5 ratings—from text and images in a single forward pass via POST /v1/decide, with a 256K token context. System 2 is the unmodified Qwen3.8-27B base for text generation and reasoning, supporting optional thinking modes ("on" or confidence-triggered "auto" below a 0.8 System 1 threshold). On the Decision Index Vision board (13,695 rows), System 1 scored 77.8% accuracy, ECE 0.019, and ~264 ms/row on an RTX PRO 6000, ranking first among 20 models with a Full score of 69.82 versus Solomon v1.1 (67.66) and Qwen3.5-397B-A17B (63.44). On ERQA (400 questions, ~5-point 95% CI), one-pass accuracy was 53.8%, rising to 60.0% with low thinking and 61.5% with medium thinking on a B200; comparisons list PhysBrain 1.5 (52.8%) and vllm-sr/d3 (58.8%). The decision head was trained on text only, making image decisions zero-shot; base pretraining data is undisclosed. Licensed Apache-2.0, the model uses Transformers, safetensors, and vLLM; no general hardware minimum or full deployment spec is provided.
Read original
dev.to