d3-ultra

d3-ultra

Answers many questions about one text, image or video input at once, with a probability for each answer and no generated text. A 27B multimodal decision model from vLLM Semantic Router.

Made by vllm-sr on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: d3-ultra is a multimodal decision model from vllm-sr, part of the decision models used by vLLM Semantic Router. You send text or JSON, and may add pictures or video clips, together with your questions. The model returns a probability for each answer instead of writing text.

Who it suits: Developers who build routers, filters or review steps that need quick, scored decisions about text, pictures or video clips.

Industry | Who uses it | How they use it | Example
Software | Platform engineers | Route or label requests by asking several questions of one input | Is this message urgent, yes or no
Media | Content reviewers | Ask where something appears in an image or a video | Find a cat in a painting
Research | ML researchers | Study a decision model that returns probabilities | Compare options with scores
Analytics | Data teams | Rate items on a scale from one shared read | Score a photo on several traits

If you want free-form writing or chat replies, you can skip it, since it does not generate text.

Where to use it:

• Windows, macOS or Linux with Python, for downloading the files
• Python programs that use the transformers library
• The free hosted API and the official Space the card points to

Use case | Who | What they get back
Choice | Developers | The probability of each option you list
Yes or no | Reviewers | A probability for each yes or no question
Box | Image and video teams | One box and a confidence that the target is there
Score | Analysts | A rating on a scale for the input

How to set it up:

• Python, to run the Hugging Face download tool
• About 52.2 GB of disk for the BF16 weights, or about 13 GB at 4-bit as an estimate
• A GPU able to serve a 27B model, since the card quotes one AMD Instinct MI325X

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Follow the card's quickstart to send your first request.

pip install -U huggingface_hub
hf download vllm-sr/d3-ultra

How to use it: You send one request holding the input and your questions. The card lists five decision types, so each question is a choice, a yes or no, a score, a box in an image or video moment, or a track over a whole video.

Tool | What it does
Choice | Picks one of several options you give
Yes / No | Returns a probability for a yes or no question
Score | Rates the input on a scale
Box | Points to a place in an image or a video moment
Track | Follows a target over a whole video

Example: You send one painting and ask ten questions about it in a single request. The card shows this with Seurat's A Sunday on La Grande Jatte. Every option comes back with its probability, and the painting is read only once.

For newcomers and experts: Start with the free hosted API the card mentions, or download the files and run the card's quickstart. Ask one yes or no question about a picture first, and read the probability it returns.

The model has 26.09B parameters, including a 0.46B vision encoder, and takes up to 131,072 tokens per question. A single read of the input serves every question. In the card's median timing on one photo, between 1 and 255 questions took 304 ms up to 2.16 s.

Send text, images or video
Add your questions
Model reads the input once
Every question gets answered
Probabilities come back

Good to know:

• License: Apache 2.0
• Cost to run: none named, though a free hosted API exists and you need your own hardware to run the files
• Data it reaches: whatever text, JSON, images and videos you send, and Hugging Face for the download
• Not for: writing or chatting, because it returns probabilities and does not generate text

On Hugging Face, 13 November 2026

• Task: zero shot classification
• Library: transformers
• Parameters: 26.1B
• Weights: about 52.2 GB in BF16; about 13 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,398
• Likes: 61
• License: apache-2.0
• Quantized versions: 8
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, feature-extraction, zero-shot-classification, decision-model, classification, system-one, multimodal, vision, video, object-detection

Links

• Model: https://huggingface.co/vllm-sr/d3-ultra
• Files: https://huggingface.co/vllm-sr/d3-ultra/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download vllm-sr/d3-ultra

Then follow the usage in its model card. Nothing in this kit runs by itself.