vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model
Summary
- The vLLM Semantic Router team has released d3-ultra, an open 27B multimodal decision model under the Apache-2.0 license.
- Unlike standard generative models, d3-ultra reads input once and returns probabilities for typed questions, including choices, yes/no, scoring, and visual grounding.
- It features 131,072 tokens of context, a Qwen3.5-derived backbone with linear and full-attention layers, and supports text, JSON, images, and video inputs.
- In private tests, d3-ultra scored 87.2 on the SR-Text benchmark, outperforming models like GLiDE V1.2 and its sibling d3-turbo.
- The model can be deployed via Hugging Face Transformers, the vLLM Semantic Router engine, or a free public API.
Why it matters
Decision models that evaluate probabilities for multiple structured questions in a single forward pass offer a high-efficiency alternative to traditional text-generating LLMs for routing, classification, and automation pipelines.