d3-edge

d3-edge

Answers choice, yes-or-no and score questions about text, images or video in one call, returning a probability for each answer. A small decision model for routing and triage engineers.

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-edge is a small decision model of about 0.59B parameters, including a 0.10B vision encoder, built on Qwen3.5-0.8B and published by vllm-sr as part of Decision 3.0 for vLLM Semantic Router. You feed it text or JSON, plus pictures or clips if you have them, along with the questions it should answer. It does not write text: it returns a probability for each possible answer, all in one call.

Who it suits: It suits engineers who build routing, moderation or triage steps and want one fast model to classify an input, and teams that already run vLLM Semantic Router.

Industry | Who uses it | How they use it | Example
Customer support | Support teams | Route a message to the right team and rate how urgent it is | Choose returns, billing or technical for a request
Retail | Operations teams | Check an uploaded photo against a question | Ask whether a receipt image lists a product
Platform engineering | Routing engineers | Add a decision step to a model router | Score a prompt before choosing which model gets it

Anyone who needs a model that writes text or chats can skip it, since it only scores answers.

Where to use it:

• Python with Transformers, loading the model by its name
• vLLM Semantic Router, whose decision models it belongs to
• Hugging Face, where an official Space offers a free try
• Windows, macOS or Linux, where the files download with Python

Use case | Who | What they get back
Route a request | Support teams | A probability for each team option
Yes or no check | Developers | A probability for yes and for no
Rate an input | Analysts | A probability for each point on a scale

How to set it up:

• Python with pip
• Transformers 5.17.0 plus torch, torchvision, pillow, opencv-python-headless, safetensors and accelerate
• Optionally flash-linear-attention, for faster GPU kernels in the linear-attention layers
• A GPU for good speed, with room for about 1.2 GB of BF16 weights

1. Install the packages with pip.
2. Optionally add flash-linear-attention.
3. Load the model with Transformers, allowing its remote code.
4. Call the model with your input and your questions.

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-edge

How to use it: Load the model, then call its system_one method with your input as the state and a set of questions. Each question has a type, its instructions and, for choices and scores, the criteria to pick from. The answers come back as probabilities.

from transformers import AutoModel

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

Question type | What it answers
Choice | Picks among the options you list
Yes / No | Says yes or no to a question
Score | Rates the input on a scale you define

Example: You send a support message about a cracked blender with a photo of the receipt. You ask which team should handle it, whether the receipt lists the item and how urgent it is. The model returns a probability for each team, for yes and no, and for each urgency level.

For newcomers and experts: Zero-shot classification means you name the categories when you ask, instead of training a classifier first. Start with a text-only example, then add an image.

Questions about one input are answered in a single call, though each runs its own forward pass over the input. Images can be PNG, JPEG or WebP, and videos MP4, WebM, MOV or MKV, read at 2 frames per second. With vLLM Semantic Router, the card measured 861 requests per second, roughly 6.1 times plain AutoModel, using a single AMD Instinct MI325X GPU.

Give text, JSON, images or video
Add your questions
One call runs the model
A probability comes back for each answer
Your app acts on the likeliest one

Title: Median response time in milliseconds
Text request: 6.6 ms
Request with an image: 58.6 ms
Request with a 10-second video: 274 ms

Title: Jev Decision Index on the official text board
Decision 2.0 0.6B: 13.6
d3-edge: 20.2

Good to know:

• Size: about 0.59B parameters, which the record puts at about 1.2 GB in BF16
• Scores: 20.2 on the Jev Decision Index 0.3.2 text board, second among entries of similar size, and 46.6 on a video test from the authors' own evaluation
• Speed: all timings were taken on a single AMD Instinct MI325X GPU, handling requests individually
• Remote code: the example loads the model with trust_remote_code, which runs code from the model repository
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the text, images and videos you pass in, while pip fetches packages from the web
• Not for: writing or chatting, since it returns probabilities and not text

On Hugging Face, 13 November 2026

• Task: zero shot classification
• Library: transformers
• Parameters: 594M
• Weights: about 1.2 GB in BF16; about 297 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,937
• Likes: 59
• License: apache-2.0
• Based on: Qwen/Qwen3.5-0.8B
• Quantized versions: 9
• Fine-tunes: 5
• 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-edge
• Files: https://huggingface.co/vllm-sr/d3-edge/tree/main

Get the model

Download its files with Hugging Face's command line:

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

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