Decision 3.0: six open, multimodal decision models, 0.6B–27B
vLLM Semantic Router ships a full Apache-2.0 family that answers in probabilities, not text — images and video in, self-hosted, weights free.
Copy markdown
Six models, 0.6B to 27B, all Apache-2.0
vLLM Semantic Router dropped the Decision 3.0 weights on Hugging Face: d3-edge (0.6B) up to d3 (27B), fine-tuned from Qwen3.5/3.8. Instead of generating text, each takes a state plus your named questions and returns a probability per option in a single forward pass — you threshold a number instead of parsing model output.
They score images and video, not just text
Five of the six read PNG/JPEG/WebP images (up to ~1.6MP each) and video at 2fps up to 32 frames, answering Choice, Yes/No and Score questions about what they see. d3 reports 71.6 on a vision board and 77.4 on Perception Test video — the first open decision family that takes pixels as input, not only tokens.
Fast enough to gate every request
On one GPU, d3 answers text in ~55ms and the 0.6B d3-edge in under 7ms; even with an image the small models stay in double-digit milliseconds. There is no hosted API — you self-host with `vllm-sr serve` on CUDA or ROCm and wire them into Semantic Router as classifiers, guardrails, or a routing layer.
Build this weekend: a local decision layer
Hand it your own schema — 'is this ticket urgent?', 'which of these five tools should the agent call?', 'does this screenshot show an error?' — and get back probabilities you can threshold, offline and free. Each question is its own forward pass, so cost scales with questions asked, not tokens generated.
The honest catch: scores are self-reported
Every benchmark is the team's own, with no published calibration metrics (Brier/ECE), so a returned 0.9 is not proven to mean 90%. Weights are BF16-only — no quantized builds yet — and downloads were still in the low hundreds a day in. Worth a weekend test on your own data before it gates anything in production.