JEV-27B-VL

The decision model that can see. JEV-27B-VL adds a vision encoder to the Jev line: a robot-arm camera frame or a browser screenshot in, a probability for every allowed action out, one forward pass at up to a 262k context. 1.53 million downloads and 2,496 likes make it the most-adopted model of the October decision wave, Apache 2.0 on the Qwen3.5 27B backbone (64 layers, GQA 4x256).

The robot and the browser are the same workload. The card’s own demos show a physical pick-and-place at about 240 milliseconds per decision and headless-Chromium computer use finishing 95 percent of 60 multi-step tasks - both are “look at pixels, score the allowed actions” before a policy layer commits. For agent stacks the practical use is action-gating: the LLM proposes, JEV-VL scores the option set in one pass, the harness acts only on high-confidence scores and asks a bigger model otherwise. 15.9GB of q4 weights plus 256MB per 1,000 tokens of cache lands it on 24GB cards for real contexts.

Where it sits. Jev (flagship reasoning) proved typed decisions at model-scale; Jeff distilled it to 2B; the October wave industrialized the shape, and the -VL variant is the one that closes the loop on agents that act in the world instead of only in text.