vLLM Semantic Router Releases Decision 3.0 Open Multimodal Decision Models

The vLLM Semantic Router team has released Decision 3.0, a family of five multimodal decision models ranging from 0.85B to 26.09B parameters under Apache-2.0. The models read text, JSON and images and return probabilities instead of generated text, simplifying classification, routing and gating. The 27B d3 tops its text and vision tables on internal evaluations, while the 9B d3-flash beats last generation's 27B Decision 2.0.

• 5 models: d3-lite (0.85B), d3-nano (2.21B), d3-mini (4.54B), d3-flash (8.39B), d3 (26.09B)
• All Apache-2.0, built on Qwen bases with vision encoders reading images up to 1.6 MP
• d3 (27B) scores 64.1 on text and 71.6 on vision in internal evals
• d3-flash (9B) scores 59.0, beating Decision 2.0's 27B at 55.9