Qwen3.6-35B-A3B

Qwen3.6-35B-A3B

Helps with coding, repository-level reasoning and image-plus-text input using a 35B mixture-of-experts model that activates 3B parameters. For developers running open weights in transformers, vLLM or SGLang.

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: Qwen3.6-35B-A3B is the first open-weight model of the Qwen3.6 line, released after the Qwen3.5 series. It is a language model with a vision encoder, tuned for agentic coding such as frontend work and reasoning across a whole repository. The card gives 35B parameters in total, with 3B activated.

Who it suits: It suits developers who want an open-weight coding model on their own hardware, and teams who compare open models using published benchmark tables.

Industry | Who uses it | How they use it | Example
Software | Developers | Run an open-weight coding model on their own hardware | Ask about a repository and its frontend code
Research | Teams comparing open models | Read the benchmark tables on the card | Compare SWE-bench Verified scores with Qwen3.5 and Gemma4
Hosting | Teams serving models | Load the files in vLLM or SGLang | Run the model behind their own tools

People with a small GPU can skip the full weights, or start from a quantized copy instead.

Where to use it:

• Hugging Face Transformers, in the format the repository holds
• vLLM, SGLang and KTransformers, which the card says the files work with
• Windows, macOS or Linux, where you can download the files with Python

Use case | Who | What they get back
Agentic coding | Developers | Help with frontend workflows and repository-level reasoning
Iterative development | Developers | Reasoning from earlier messages kept in context
Image plus text input | Teams building assistants | Answers to prompts that include pictures

How to set it up: You need these first:

• Python
• The huggingface_hub package
• Room for the weights, about 71.9 GB in BF16 or about 18 GB at 4-bit as an estimate
• One of the engines the card names, such as Transformers or vLLM

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Load them in Transformers, vLLM, SGLang or KTransformers.

pip install -U huggingface_hub
hf download Qwen/Qwen3.6-35B-A3B

How to use it: Download the files, load them in one of the engines the card names, then chat with the model, with pictures if you like. The card text here has no run commands, so follow its model card for the exact calls. The card also describes a Thinking Preservation option that keeps reasoning from earlier messages.

Example: A person could ask it to find why a frontend component breaks inside a repository, and turn on the option that keeps reasoning from earlier messages. The reply would carry its reasoning forward as the work goes on.

For newcomers and experts: This is a large model built for coding help that can also read images. The files are big, so a first small step is to look at a quantized copy, since Hugging Face lists 974 of them from other people.

Under the hood it is a mixture of experts with 256 experts, 8 routed plus 1 shared active at a time. It has 40 layers laid out as ten repeats of three Gated DeltaNet layers and one Gated Attention layer, each followed by a mixture-of-experts block. Its context window is 262,144 tokens by default, and it can stretch to 1,010,000.

Download the weights with hf
Load them in a supported engine
Send text or an image with a prompt
Optionally keep earlier reasoning
Read the reply

Good to know:

• Size: about 71.9 GB in BF16 and about 18 GB at 4-bit, as an estimate
• Versions: Hugging Face lists 974 quantized and 439 fine-tuned versions made by other people
• Benchmarks: the card reports scores such as SWE-bench Verified, compared with Qwen3.5 and Gemma4 models

Title: SWE-bench Verified scores
Qwen3.5-27B: 75.0
Qwen3.6-35BA3B: 73.4
Qwen3.5-35BA3B: 70.0
Gemma4-31B: 52.0
Gemma4-26BA4B: 17.4

• License: apache-2.0
• Cost to run: none named, no API key or account
• Data it reaches: it downloads model files from Hugging Face, and the card names no other service or account

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: 4,071,952
• Likes: 3,119
• License: apache-2.0
• Quantized versions: 981
• Fine-tunes: 469
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen3.6-35B-A3B
• Files: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3.6-35B-A3B

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