Qwen3.5-35B-A3B

Qwen3.5-35B-A3B

A multimodal chat model that reads images and text, with 35B parameters of which 3B are active at a time. For developers who host their own large model with Transformers, vLLM or similar.

Made by Qwen 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: Qwen3.5-35B-A3B is a post-trained model from Qwen that takes images and text and answers in text. It uses a mixture-of-experts design with 35B parameters in total and 3B active per token. The repository holds weights and configuration files in the Hugging Face Transformers format.

Who it suits: Developers and teams who want to run a large open-weight model themselves, including one that reads images.

Industry | Who uses it | How they use it | Example
Software | Developers | Host their own chat and coding model | A coding helper behind an internal tool
Research | AI researchers | Study a mixture-of-experts model with vision | Compare it with other open models
Enterprise | Platform teams | Serve a model on their own servers | An internal assistant that reads screenshots
Education | Language teams | Use its multilingual coverage, listed at 201 languages and dialects | A tutor that answers in several languages

You can skip it if you have no server-class GPU memory, or if you only want a hosted chat. The maker points to a hosted version for that.

Where to use it:

• Hugging Face Transformers
• vLLM, SGLang and KTransformers, which the card names as compatible
• Windows, macOS or Linux, with Python for the download step
• Alibaba Cloud Model Studio, where the hosted Qwen3.5-Flash version runs

Use case | Who | What they get back
Image and text questions | Developers | Text answers about a picture you supply
Self-hosted assistant | Platform teams | A model on your own hardware
Long documents | Analysts | Context of 262,144 tokens, stretchable to 1,010,000

How to set it up:

• Python and pip on your computer
• Large GPU memory, since the BF16 weights are about 71.9 GB and a 4-bit copy is about 18 GB by estimate
• One of the compatible runtimes the card names

1. Install the Hugging Face hub package.
2. Download the model files with the hf command.
3. Open the model card and your chosen runtime's docs to serve the model.

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-35B-A3B

How to use it: Load the files in Transformers or one of the other named runtimes, then send it an image, text or both. The card describes the repository as the post-trained model, ready for those tools.

hf download Qwen/Qwen3.5-35B-A3B

Example: You send a screenshot of an error dialog and ask what it means. The model reads the image and answers in text with an explanation and likely causes.

For newcomers and experts: If you are new, a model this size needs a lot of memory, so start by reading the model card and checking your hardware. If you cannot host it, the hosted Qwen3.5-Flash is the easier path.

If you know the field, the card describes 40 layers laid out as repeating Gated DeltaNet and Gated Attention blocks, each followed by mixture-of-experts. There are 256 experts, with 8 routed and 1 shared active. Context is 262,144 tokens natively. The weights come from the Qwen3.5-35B-A3B-Base checkpoint.

Image and text input
Vision encoder plus language model
Expert layers, 3B active parameters
Text answer

Title: Weight size in GB
BF16: 71.9 GB
4-bit estimate: 18 GB

Good to know:

• License: Apache 2.0
• Cost to run: your own GPU hardware, or the hosted Qwen3.5-Flash on Alibaba Cloud Model Studio, whose price the card does not give
• Data it reaches: the files you download and the images and text you send it
• Versions: others have published 317 quantized and 202 fine-tuned copies

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 36B
• Weights: about 71.9 GB in BF16; about 18 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,514,731
• Likes: 1,559
• License: apache-2.0
• Based on: Qwen/Qwen3.5-35B-A3B-Base
• Quantized versions: 318
• Fine-tunes: 204
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen3.5-35B-A3B
• Files: https://huggingface.co/Qwen/Qwen3.5-35B-A3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-35B-A3B

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