vLLM Semantic Router Releases d3-ultra, a 27B Open Multimodal Decision Model
vLLM Semantic Router released d3-ultra, a 27B Apache-2.0 multimodal decision model that returns probabilities, not text.
d3-ultra has 26.09 billion dense parameters, including a 0.46 billion-parameter vision encoder, and supports 131,072 tokens of context per question. It was tested on one AMD Instinct MI325X with BF16 weights; no official quantized variant is listed yet. The model is the newest member of the Decision 3.0 family, which spans 0.6 billion to 27 billion parameters.
The model reads an input once per request and answers every question from that single read, a design the team calls “state once.” A user sends a state, in text or JSON plus images or videos, and a set of typed questions. Question types include Choice, Yes/No, Score, Box and Track. Images can also appear inside questions and options. Box answers always return exactly one box with a confidence; a confidence near zero means the target is not in the picture. Questions over 131,072 tokens fail with max_length_exceeded, and nothing is silently truncated.
Its configuration shows a Qwen3.5-architecture backbone with 64 layers, 48 of them linear-attention layers and 16 full-attention layers. The model card does not name the base checkpoint. The vLLM recipe says the sibling d3 was trained from Qwen3.8-27B.
In the team’s own private tests, d3-ultra scored 87.2 on SR-Text, the highest among six decision models. GLiDE V1.2 scored 85.1, d1-ultra 83.4, d3-turbo 83.1, d3 81.4 and Jev 1.14 73.1. SR-Text contains 10,411 text requests with 73,495 questions, and scores are chance-corrected. d3-ultra led every area except arts and human taste, where d1-ultra scored 90.8 against d3-ultra’s 89.6, a difference within the margin of error. The model’s best result was 97.8 on SR-Text tools and automation.
On SR-Box, which contains 5,678 box questions on 4,984 images across 10 domains and counts a box as correct at IoU 0.5 or higher, d3-ultra scored 58.2% across 10 domains and 54.6% across nine. d3-turbo scored 60.0% and 59.6%, while Jev 1.14 scored 45.7% and 46.9%. The team says d3-turbo remains stronger on natural images, while d3-ultra leads on charts. All benchmark numbers are self-reported on private test sets.
On latency, one question about a photo took 304 ms on one AMD Instinct MI325X with AutoModel, according to MarkTechPost. A batch of 255 questions took 2.16 seconds, about 7.3 ms per additional question.
The model can be deployed through Transformers using AutoModel.from_pretrained with trust_remote_code and transformers 5.17.0; an optional flash-linear-attention package adds fast GPU kernels. It can also run through the vLLM Semantic Router Engine with the command vllm-sr serve -e vllm-sr/d3-ultra, added in PR #5449 on Oct. 25, 2026. The model card says a free API is available at decision.vllm-sr.ai, and a demo is hosted on Hugging Face Spaces.