vLLM Semantic Router releases Decision 3.0: six open sizes that read text, images and video

Source · Model card: Decision 3.0 (d3)

On 10 October 2026 vLLM Semantic Router published Decision 3.0 (announcement): six Apache-2.0 sizes, d3 (26B, Qwen3.8-27B), d3-flash (8.4B), d3-mini (4.5B), d3-nano (2.2B), d3-lite (0.85B) and d3-edge (0.59B). All read text or JSON plus images and video and answer Choice, Yes/No and Score questions.

Vendor numbers (UnverifiedClaim): Decision Index 0.3 public suite with the official kit: 65.45, 57.79, 54.90, 42.22, 36.93 and 28.82 from largest to smallest. The 27B card also claims a full score of 64.7, which would be #1, but the entry was not in the board data we read. Median text latency 55 ms (27B) to 6.7 ms (edge) on an AMD MI325X.

New catalog page: Decision 3.0, 6.5 / 7.6 / 3.