d3-flash

d3-flash

A 9B model that answers several choice, yes or no and score questions about text, images or video in one call, returning a probability for every answer instead of generated text. For routing and decision tasks.

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-flash is the 9B multimodal decision model of Decision 3.0, built for vLLM Semantic Router. You hand it an input and a set of questions, and it answers all of them in one call. Each answer comes back as a probability for every option, and no text is generated.

Who it suits: Developers who need to sort, check or rate inputs automatically, such as incoming messages or documents with images. It fits systems that want a fast, scored decision rather than a written reply.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Route each request to the right team and flag urgency | Score a damaged-delivery message for team, receipt and urgency
E-commerce | Order-handling teams | Check an attached receipt image against a claim | Ask whether the receipt lists the product
AI infrastructure | Teams running vLLM Semantic Router | Use it as the decision model in a routing setup | Send each request's questions to the model and read the probabilities
Operations | Workflow builders | Turn free-form input into choices, yes or no answers and ratings | Rate how urgent an incoming request is

You can skip it if you want written answers, chat or summaries. It returns probabilities, not text.

Where to use it:

• Python on Windows, macOS and Linux, through the transformers library
• vLLM Semantic Router, which the card names as its serving setup
• GPU servers, where the card measured speed on a single AMD Instinct MI325X

Use case | Who | What they get back
Routing a request | Support and ops teams | A probability for each team you list
Yes or no checks | Developers | A probability for yes and for no
Rating on a scale | Workflow builders | A probability for each level you define
Questions about images or video | Product teams | Answers that take attached images or clips into account

How to set it up: You need Python and the packages below. A GPU speeds it up.

• Python on Windows, macOS or Linux
• The pip packages from the quickstart
• Optionally, flash-linear-attention for fast GPU kernels

1. Install Python if you do not have it.
2. Install the packages with the commands below.
3. Load the model with transformers in your script.

pip install "transformers==5.17.0" torch torchvision pillow opencv-python-headless safetensors accelerate
pip install flash-linear-attention  # optional: fast GPU kernels for the linear-attention layers

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

How to use it: Load the model, then call system_one with the input as state and a questions dictionary. Each question has a type, instructions and criteria. The result holds the answers with their probabilities.

import json

from huggingface_hub import hf_hub_download
from transformers import AutoModel

model = AutoModel.from_pretrained("vllm-sr/d3-flash", trust_remote_code=True)

# Text
result = model.system_one(
    state="The order arrived damaged yesterday. The customer has a receipt and asks for a replacement today.",
    questions={
        "route": {
            "type": "choice",
            "instructions": "Which team should handle this request?",
            "criteria": {
                "returns": "Refunds, replacements and damaged deliveries",
                "billing": "Payments, invoices and charges",
                "technical": "Product setup and faults"
            }
        },
        "receipt": {
            "type": "noul",
            "instructions": "Does the customer have a receipt?"
        },
        "urgency": {
            "type": "score",
            "instructions": "How urgent is this request?",
            "criteria": [
                "Routine",
                "Soon",
                "Today"
            ]
        }
    },
)
print(json.dumps(result["answers"], indent=2))

For images, add an images argument holding file paths, web links, PIL objects or base64 strings. Videos accept file paths, links, base64 strings or arrays of frames.

Tool | What it does
system_one | Answers all your questions about one input in a single call
choice question | Picks among options you describe in criteria
yes or no question | Returns a probability for each side
score question | Rates the input on the levels you list

Example: You send a message saying a delivered order arrived damaged and the customer has a receipt, with three questions. One asks which team should handle it, one asks about the receipt and one rates urgency. Back come probabilities for every option of each question.

For newcomers and experts: Think of it as a quick sorter. You give it something to look at and a few questions with fixed answers, and it tells you how likely each answer is. Start by running the text example and reading the printed answers.

The model is 8.39B parameters, including a 0.46B vision encoder. Every question gets its own pass through the model over the input, which is why no text generation is needed. Timings on a single AMD Instinct MI325X: 19.2 ms median for text only, 166 ms when an image is attached, 371 ms when a 10-second clip is attached. Videos are read at 2 frames per second.

Send text, JSON, images or video
Add choice, yes or no and score questions
One call runs a pass per question
Get a probability for every answer

Title: Median latency on one AMD Instinct MI325X, in milliseconds
Text request: 19.2 ms
Request with an image: 166 ms
Request with a 10-second video: 371 ms

Good to know:

• Images can be PNG, JPEG or WebP, and videos can be MP4, WebM, MOV or MKV, several per request.
• Loading uses trust_remote_code, which runs code that ships with the model, so read it first.
• With vLLM Semantic Router the card reports 95 requests per second on one MI325X.
• License: Apache-2.0.
• Cost to run: none named. The card lists a free demo and uses an AMD GPU for its own timings.
• Data it reaches: files from Hugging Face, plus any text, images, videos or URLs you pass in.
• Not for: writing text, since it returns probabilities without generating text.

On Hugging Face, 13 November 2026

• Task: zero shot classification
• Library: transformers
• Parameters: 8.4B
• Weights: about 16.8 GB in BF16; about 4.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 103,577
• Likes: 114
• License: apache-2.0
• Based on: Qwen/Qwen3.5-9B
• Quantized versions: 17
• Fine-tunes: 15
• Tags: transformers, safetensors, qwen3_5, feature-extraction, zero-shot-classification, decision-model, classification, system-one, multimodal, vision, video

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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