vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model

The vLLM Semantic Router team released d3-ultra under Apache-2.0. It is a decision model: you hand it a state, which can be text, JSON, images or video, plus a list of questions, and it returns probabilities for every option. Question types include multiple choice, yes/no, scores, bounding boxes in an image, and tracking a target through a video.

The pitch is efficiency. The model reads the input once and answers every question from that single pass. On one AMD Instinct MI325X, the team reports one question about a photo in 304 milliseconds and 255 questions in 2.16 seconds. For anyone routing requests, classifying tickets or checking images at volume, that is a different cost curve than asking a generative model to reason in prose every time.

We think this is the unglamorous plumbing that makes AI useful. A lot of real deployments are not conversations; they are thousands of small judgments. Which model should handle this request? Is this ticket billing or support? Is the product actually in the photo? A model built to answer those questions directly, with a calibrated probability instead of a paragraph, is a sensible design, and putting it out with open weights under a permissive license makes it easy to inspect and self-host.

Some details we appreciate. Box answers always return exactly one box with a confidence, and a near-zero confidence signals the target is absent. Inputs that exceed the 131,072-token limit fail with an explicit error rather than being silently truncated. That is the kind of honest failure mode production systems need.

Now the caution. Every benchmark number here is self-reported on the team's own private test sets. d3-ultra's 87.2 on SR-Text, ahead of Fastino's GLiDE V1.2 at 85.1, comes from a test the publisher designed, and its public index score is submitted but not yet on the board. It also loses to its own sibling, d3-turbo, on image grounding, 58.2 percent to 60.0 across ten domains, with a wider gap on nine. And it was tested on a single high-end AMD accelerator in BF16 with no official quantized version yet, which puts self-hosting out of reach for many smaller teams.

The bigger risk is what happens when probability-scoring models quietly become the gatekeepers. Ticket routing and content checks sound harmless until they decide which customer gets a human, which claim gets flagged or which image gets removed. A clean number between zero and one can look more objective than it is. If this pattern scales, we want those decisions logged, audited and contestable.

Our read: promising, well-documented, and worth testing for routing and classification. Wait for independent evaluations before treating the leaderboard as settled.

d3-ultra is the largest member of a Decision 3.0 family ranging from 0.6B to 27B parameters. Its configuration shows a Qwen3.5-architecture backbone with 64 layers, 48 of them linear attention; the model card does not name its base checkpoint. It can be run through Transformers, the vLLM Semantic Router engine, a free hosted API, or a Hugging Face demo.