The vLLM Semantic Router team released d3-ultra on October 26, 2026, introducing a 27B open multimodal decision model under the Apache-2.0 license. Unlike a text-generation model, d3-ultra does not generate text. It reads an input once, then returns a probability for every answer to every typed question.
That design gives developers one open-weight call for routing, classification, visual grounding, and video checks. A quiz could show one photo and offer four etchings as possible answers, while a business system could use the same model for ticket routing or SQL checks.
The central idea is simple: state once. The model reads the input once per request and answers every question from that single read. The model has 26.09B dense parameters, including a 0.46B vision encoder, and supports 131,072 tokens of context per question.
A decision model built for many questions
d3-ultra is the newest member of Decision 3.0, a family that spans 0.6B to 27B. Its configuration shows a Qwen3.5-architecture backbone with 64 layers: 48 linear-attention layers and 16 full-attention layers. The model card does not name its base checkpoint, though vLLM’s recipe says the sibling d3 was trained from Qwen3.8-27B.
For box questions, the model always returns exactly one box with a confidence value. A confidence near 0 means the target is not in the picture. Questions longer than 131,072 tokens fail with max_length_exceeded, so the system does not silently truncate the input.
The team tested d3-ultra on one AMD Instinct MI325X with BF16 weights. No official quantized variant is listed yet. Developers have four ways to try it: Transformers, the vLLM Semantic Router Engine, a free API, and the official Hugging Face Space demo.
The Transformers route uses AutoModel with transformers==5.17.0 and the model identifier vllm-sr/d3-ultra. The optional flash-linear-attention package adds fast GPU kernels. The engine can serve the model with vllm-sr serve -e vllm-sr/d3-ultra, a route added in PR #5449 on October 25, 2026. The model card also says decision.vllm-sr.ai serves d3-ultra with a free key.
Strong text results, weaker image grounding
The research team compared d3-ultra with GLiDE V1.2, d1-ultra, d3-turbo, d3, and Jev 1.14 across two in-house tests. Scores use chance correction, where 0 means chance and 100 means perfect.
On SR-Text, d3-ultra recorded the highest score among the six decision models in the team’s private test. Its overall score was 87.2, ahead of GLiDE V1.2 at 85.1. Other listed comparison scores include 83.4, 83.1, 81.4, and 73.1.
SR-Text contains 10,411 text requests and 73,495 questions. Most of those questions recreate skills from 34 public benchmarks with new items, while another 4,896 questions cover 43 new task families, including SQL checks and ticket routing.
d3-ultra leads every SR-Text area except arts and human taste. In that area, d1-ultra scores 90.8 against d3-ultra’s 89.6, a difference within the margin of error. The strongest listed d3-ultra result is 97.8 on tools and automation.
The picture changes on SR-Box. That test contains 5,678 box questions across 4,984 images and 10 domains, with a box counted as correct at IoU 0.5 or higher. d3-ultra scores 58.2% across 10 domains and 54.6% across 9, while d3-turbo reaches 60.0% and 59.6%.
That puts d3-ultra 5.0 points below d3-turbo on the 9-domain result. Jev 1.14 scores 45.7% and 46.9% on the same measures. The team says d3-turbo remains stronger on natural images, while d3-ultra leads on charts.
Many questions without repeating the image read
The latency results show why the one-pass approach matters. On one AMD Instinct MI325X using AutoModel, one question about a photo takes 304 milliseconds. Asking 255 questions about one photo takes 2.16 seconds, or about 7.3 milliseconds for each extra question.
That pattern makes d3-ultra suited to tasks where one image, video, or other input needs many decisions. Instead of sending the same input through separate calls for each question, developers can ask the full set in one request and receive probabilities for every answer.
d3-ultra is not the strongest choice for every visual task. Its SR-Box score trails d3-turbo, especially on natural images, but its broader text performance and chart results show where the model is aimed: open multimodal decisions, not generated prose.
Based on
- vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model — marktechpost.com
- Physical AI: When AI Stops Generating And Starts Acting On The World — forbes.com
- Riemann Computing Inc. Creates and Demos Out a 13B Parameter AI Model, Aims at Massively Reducing Training Costs | AP News — apnews.com
- DEEP Robotics Unveils Five Research Advances in Embodied AI | Markets Insider — markets.businessinsider.com
- What If AI’s Black Box Isn’t As Mysterious As We Thought? — forbes.com
