Vllm semantic router launches d3-ultra: a 27b multimodal decision model

The vLLM Semantic Router team has unveiled d3-ultra, a 27B-parameter multimodal decision model under the Apache-2.0 license. This model does not generate text but instead assigns probabilities to answers for every question. It is part of the Decision 3.0 family, spanning from 0.6B to 27B parameters. d3-ultra processes input once per request and answers all questions from that single read. It uses a Qwen3.5 architecture with 64 layers, including 48 linear-attention and 16 full-attention layers. The model excels in tasks like image and video checks, routing, and classification. In benchmark tests, d3-ultra outperformed other models in most categories, except for arts and human taste, where d1-ultra slightly led. The model handles up to 131,072 tokens, with a latency of 304 ms per question on an AMD Instinct MI325X. It is available via a model card, GitHub repo, and Hugging Face Space, with additional engagement options on Twitter, Reddit, and Telegram.