AutoTrust open-sources JEV-27B-VL: the first open-weight near-SOTA multimodal decision model — it watches Mario and decides whether to jump
AutoTrust AI (Singapore, released Sep 29) open-sources JEV-27B-VL, billed as the world's first open-weight, near-SOTA multimodal decision model built with its Blocks of Experts (BoE) recipe: a frozen Qwen3.8-27B backbone serves as one expert block; a small detachable decision block of 108.9M parameters (~0.4% of the model, ~9.2 h on one B200) is trained on top; a router dispatches each request to either the fast decision block or the deliberate generation block, serving System 1 and System 2 from one set of weights. It doesn't just describe what it sees — it converts visual state into **calibrated action probabilities** and then picks the next move: in the demo it watches Mario's position, motion, obstacles and timing, and chooses whether to run, jump, or both (See → decide → act). The text-only JEV-27B averages 84.07% across six public decision benchmark groups (JevBench 88.70, Nimble 92.91, MASSIVE-en 87.71, etc.), self-measured slightly above its distillation target TypeSafe Jev 1.13 (83.85%); on 25,376 held-out questions labeled by Jev itself the mean KL divergence is only 0.017. Trained on a public Apache-2.0 corpus of Jev 1.13 output distributions compiled via OpenRouter. Apache-2.0 weights on HuggingFace (autotrust/JEV-27B-VL); self-hostable on a single B200 via vLLM or Docker.
