vLLM Semantic Router Releases Decision 3.0

The vLLM Semantic Router project has released Decision 3.0, a suite of five Apache-2.0 open multimodal decision models. Built by fine-tuning Qwen base models, the collection includes d3-lite (0.85B parameters with a 0.10B vision encoder), d3-nano (2.21B with 0.33B vision), d3-mini (4.54B with 0.33B vision), d3-flash (8.39B with 0.46B vision), and d3 (26.09B based on Qwen3.8-27B with 0.46B vision). These models take text, JSON, and PNG, JPEG, or WebP images up to 1.6 megapixels, returning output probabilities for predefined choices, binary questions, or scaled scores rather than generating standard text.

Evaluated on an AMD Instinct MI325X GPU in BF16 precision, single-request median text and image latencies reach 23.5 ms and 60.7 ms for d3-lite, 21.7 ms and 92.0 ms for d3-nano, 28.5 ms and 115.0 ms for d3-mini, and 29.7 ms and 171.8 ms for d3-flash. The 27B d3 variant logs a median text latency of 84 ms. Deployment requires transformers==5.17.0 with trust_remote_code=True, with optional flash-linear-attention support.

Internal tests on the Jev Decision Index 0.3.1 place d3 top in text (64.1) and vision (71.6), outpacing Perplexity Decider v1.1's 62.8 and 70.6. On public suites, Torchcast Decision 27B edges d3 (65.1 to 64.9). The 9B d3-flash scores 59.0, topping Decision 2.0's 27B model (55.9). Public suite gains over Decision 2.0 range from +8.0 to +15.1. On specific tasks, d3 reached 97.7 on KIE (CORD+FUNSD) document extraction, 97.3 on InfographicVQA, and 89.9 on Mind2Web, but scored 46.2 on MMMU-Pro vision and 44.6 on Hateful Memes moderation.

Among smaller sizes, d3-mini and d3-flash top text and vision tables for their classes. Conversely, d3-lite ranks third in vision (41.1, behind JPT-0.8B's 42.1), and d3-nano trails LiquidAI d1-3B on text (35.8 versus 40.1). For engineers, Decision 3.0 provides calibrated scoring for request routing and gating without LLM text parsing.