The vLLM Semantic Router project has released d3-ultra, a 27 billion parameter multimodal decision model licensed under Apache-2.0. This model does not generate text. It accepts an input once and returns a probability for every possible answer to every question. For developers, this enables routing, classification, visual grounding, and video verification from a single open-weight call.
Size and hardware
The model contains 26.09 billion parameters in dense form, including a 0.46 billion parameter vision encoder. It processes 131,072 tokens of context per question. Testing was performed on a single AMD Instinct MI325X using BF16 weights. No official quantized variant is currently listed.
In the team’s private tests, d3-ultra achieved the highest SR-Text score among six decision models. It scored 97.8 on SR-Text tools and automation, measured as chance-corrected accuracy. On SR-Box, it scored 54.6% across nine domains, which is 5.0 points below its sibling model, d3-turbo.
What is d3-ultra?
d3-ultra is a System One decision model from the vLLM Semantic Router suite. It is the newest addition to the Decision 3.0 family, which ranges from 0.6 billion to 27 billion parameters. Users send a state, which can be text or JSON, along with images or videos, and a set of questions.
Each question falls into one of five types:
- Choice: select one option from several.
- Yes / No: return the probability of a yes answer.
- Score: rate an item on a scale.
- Box: locate an object within an image or video moment.
- Track: follow a target across an entire video.
Images may appear inside questions and options. A quiz could display a photo and then offer four etchings as potential answers.
Architecture and operation
The core concept is state once. The model reads the input once per request and answers every question from that single pass.
The configuration shows a Qwen3.5-architecture backbone with 64 layers. Forty-eight layers use linear attention, while 16 use full attention. The card does not name the base checkpoint. The vLLM recipe states the sibling d3 model was trained from Qwen3.8-27B.
Box answers always return exactly one box with a confidence score. A confidence near zero indicates the target is not in the picture. Questions exceeding 131,072 tokens fail with a max_length_exceeded error; nothing is silently truncated.
Benchmark performance
The research team compares d3-ultra against GLiDE V1.2, d1-ultra, d3, Jev 1.14, and d3-turbo on two in-house tests.
SR-Text consists of 10,411 text requests and 73,495 questions. Most recreate skills from 34 public benchmarks with new items. Another 4,896 questions come from 43 new task families, such as SQL checks and ticket routing. Scores are chance-corrected, where 0 is random chance and 100 is perfect.
- d3-ultra: 87.2
- GLiDE V1.2: 85.1
- d1-ultra: 83.4
- d3-turbo: 83.1
- d3: 81.4
- Jev 1.14: 73.1
d3-ultra leads every area except arts and human taste. In that category, d1-ultra scores 90.8 against d3-ultra’s 89.6, a difference within the margin of error.
SR-Box includes 5,678 Box questions on 4,984 images across 10 domains. A box counts as correct at an Intersection over Union (IoU) of 0.5 or higher. d3-ultra scores 58.2% across 10 domains and 54.6% across 9. d3-turbo scores 60.0% and 59.6%. Jev 1.14 scores 45.7% and 46.9%. The team notes d3-turbo remains stronger on natural images, while d3-ultra leads on charts.
Latency: on a single AMD Instinct MI325X with AutoModel, one question about a photo takes 304 ms. Twenty-five-five questions take 2.16 seconds, averaging about 7.3 ms per extra question.
Deployment options
There are four routes for deployment:
- Transformers: Use
AutoModel.from_pretrained("vllm-sr/d3-ultra", trust_remote_code=True)withtransformers==5.17.0. The optionalflash-linear-attentionpackage adds fast GPU kernels. - vLLM Semantic Router Engine: Run
vllm-sr serve -e vllm-sr/d3-ultra. This was added in PR #5449 on October 25, 2026. - Free API: The card states decision.vllm-sr.ai serves d3-ultra with a free key.
- Demo: The official Hugging Face Space is available at huggingface.co/spaces/vllm-sr/d3-ultra.
Comparison table
Below is a summary of d3-ultra against GLiDE V1.2, d1-ultra, Jev 1.14, and d3-turbo.
| Feature | d3-ultra | d3-turbo | GLiDE V1.2 | d1-ultra | Jev 1.14 |
|---|---|---|---|---|---|
| Developer | vLLM Semantic Router | vLLM Semantic Router | Fastino Labs | Liquid AI | TypeSafe |
| Parameters | 26.09B | 26.09B | 26B | Not disclosed | Not disclosed |
| Context per question | 131,072 | 32,768 default; 131,072 with long_context | 40,000 default | 32,768 | 262,144 |
| Weights / license | Open, Apache-2.0 | Open, Apache-2.0 | Open, Fastino Community License | Hosted only | Hosted only, proprietary |
| Inputs | Text, JSON, images, video | Text, JSON, images, video | Text, JSON, images | Text, images | Text, images |
| Box / track | Box + track | Box + track | No | No | Box only |
| SR-Text (self-reported by vLLM-SR) | 87.2 | 83.1 | 85.1 | 83.4 | 73.1 |
| SR-Box, 10 domains | 58.2% | 60.0% | Not supported | Not supported | 45.7% |
| Public index score | Submitted, not yet on board | 70.2 Jev DI (own estimate) | 71.5 Decision Index 0.4.1 | 65 AA Decision Index | 57 AA Decision Index |
| Price | Free API, self-host | Free API, self-host | $0.04 / M input | $0.042 / M input | $0.042 / M input |
What it means
For people building applications, d3-ultra offers a way to handle complex multimodal queries without generating text responses. This reduces the load on separate generation models. The ability to answer 255 questions about a single photo in under 2.2 seconds makes it viable for real-time interfaces. However, users should note that the model trails d3-turbo on visual grounding tasks, and all benchmark numbers are based on private test sets reported by the team.