New AI Models Make Decisions From Text, Images, and Video

With the vLLM Semantic Router team, Decision 3.0 (d3) is out: a new SOTA decision model family ranging from 0.6B to 27B, Apache-2.0, all reading text, images and video. Built on Qwen3.5 (0.8B / 2B / 4B / 9B) and Qwen3.8-27B, with parameter counts that include the vision encoder (0.33B in the 4B, 0.46B in the 27B). They handle three decision types — choice, yes/no, score — giving a probability for every answer with no text generation, and several questions about the same input are answered in one call, each from its own forward pass over the input.

Images: multiple per request, read at up to 1.6 MP. Video: multiple per request (MP4, WebM, MOV, MKV), sampled at 2 fps, at most 32 frames over the whole clip, each at up to 0.2 MP. BF16 weights, loaded with Transformers (trust_remote_code=True), trained on AMD Instinct MI325X.

In the team's evaluation with the official Jev 0.3 kit on the released weights, the 27B is #1 on both boards, scoring 64.7 text and 71.6 vision, ahead of every listed model, Jev's own included — and every size (0.6B, 0.8B, 4B and 9B) leads its size class, with the 9B (d3-flash) beating the previous-generation 27B (Decision 2.0). Video was zero-shot: no model saw video in training. On all 19,140 Perception Test validation questions, the 27B scores 77.4 and the 4B scores 70.3.