Evidence at a glance
Not a Chat Model, but an Executor of Judgments
The vLLM Semantic Router team has released d3-ultra, an open-weight multimodal decision model with roughly 27 billion parameters under the Apache-2.0 license. It is aimed not at open-ended conversation but at tasks such as classification, routing, and visual checks that need machine-consumable decisions. A caller submits text or JSON, optionally with images or video, and a set of questions. The model returns probabilities, scores, boxes, or tracks rather than a natural-language reply.
That interface changes the work the surrounding system must do. In a typical generative flow, a model writes an answer and the application then parses it and checks whether its format is valid. d3-ultra makes question types and output forms explicit in the request, so results can feed more directly into downstream code. Its value depends not only on whether a judgment is correct, but also on whether one input can serve many judgments efficiently.
One Read Pays Off as the Question Count Grows
The central mechanism is to let multiple questions in one request share the processing of the same input state. A developer can ask several choice, yes-or-no, or localization questions about one image without reprocessing it for each question or extracting conclusions from generated prose. The model supports question types including Choice, Yes/No, Score, Box, and Track. A Box answer returns one box with a confidence value, while Track is intended to follow a target through video.
The team's latency figures illustrate where this design may fit. In a test using AutoModel on one AMD Instinct MI325X, one question about a photo took 304 milliseconds, while 255 questions took 2.16 seconds in total. The team reports an increase of about 7.3 milliseconds for each question after the first. That suggests high-fan-out tasks may spread the cost of reading an input across many judgments, but the figures come from a specific hardware and test setup, not a general latency guarantee for other devices or production workloads.
A Text Lead Does Not Mean a Lead on Every Visual Task
On the team's in-house SR-Text test, d3-ultra scored 87.2, ahead of GLiDE V1.2 at 85.1, d1-ultra at 83.4, and d3-turbo at 83.1. The test contains 10,411 requests and 73,495 questions. The team says most items were newly written to exercise capabilities drawn from 34 public benchmarks, alongside additional task families. Its score of 97.8 on tools and automation makes structured decision tasks a promising area for closer evaluation.
These figures are not results from a public, standardized leaderboard: both SR-Text and SR-Box are private tests built by the team, so outsiders cannot inspect the full question sets or reproduce the scores independently. More importantly, d3-ultra scored 58.2% across ten SR-Box domains, below d3-turbo's 60.0%. Excluding charts, the scores were 54.6% and 59.6%, respectively. The team says d3-turbo is stronger on natural images while d3-ultra leads on charts. A single aggregate score therefore cannot replace evaluation against the specific tasks a business needs.
Open Weights Do Not Mean Low-Cost Deployment
The model has 26.09 billion parameters, including a 0.46-billion-parameter vision encoder, and the model card sets a limit of 131,072 tokens per question. Its configuration uses a Qwen3.5 architecture with 64 layers: 48 linear-attention layers and 16 full-attention layers. The model card does not disclose the base checkpoint. The team's reported runtime test used BF16 weights on a single MI325X, and the supplied materials list no official quantized variant. Apache-2.0 provides licensing openness, but it does not remove the memory, hardware, or operational requirements.
There are several deployment routes for different integration needs. With Transformers, the example loads the model through AutoModel and calls for transformers 5.17.0, with flash-linear-attention available as an optional install. The team also provides a vLLM Semantic Router Engine serving path, recorded in PR #5449 on October 25, 2026, and the model card lists a free API. Choosing local deployment or an API is not merely an engineering preference; it affects resource costs, service boundaries, and integration. A question exceeding 131,072 tokens returns max_length_exceeded, and the materials say the input is not silently truncated.
Evaluate Video Support Alongside Its Sampling Limits
Video support does not imply continuous, lossless understanding of an entire clip. The runtime documentation says video is sampled into frames according to configuration and is subject to limits on both frame count and video input tokens. What Track and related capabilities can actually observe therefore depends on the sampling strategy and input limits. The available materials do not establish full frame-by-frame analysis, so video use cases should be tested specifically for whether important actions are missed during sampling.
For a technical lead, a sensible pilot is not to “replace the general-purpose multimodal model.” Instead, select a workflow in which one input must produce many structured results, such as several checks on a single image or multiple routing decisions from one piece of content, and compare the end-to-end cost with the existing approach. Evaluate task accuracy, failure handling, latency, and hardware use together, treating private benchmark scores as a vendor-provided signal rather than a guarantee of business outcomes. If the workload depends mainly on natural-image localization, fine-grained video coverage, or low-cost deployment, the current evidence is not enough to assume d3-ultra is the better fit.