Qwen3.6-35B-A3B-FP8

Qwen3.6-35B-A3B-FP8

Reads text and images and writes text, as an FP8 copy of Qwen3.6-35B-A3B with 3B active parameters. For developers who run coding agents and want a smaller download than full precision.

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: This is the FP8-quantized copy of Qwen3.6-35B-A3B, the first open-weight Qwen3.6 model. It is a mixture-of-experts language model with a vision encoder, 35B parameters in total and 3B active. The quantization uses fine-grained fp8 with block size 128, and the card says its metrics are nearly identical to the original.

Who it suits: Developers who want an open model for agentic coding, with a smaller footprint than the full-precision weights.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a coding agent on a local or hosted model | Let an agent reason across a repository
Web development | Frontend teams | Use it for frontend workflows | Build a page component with agent help
AI platforms | Infrastructure teams | Serve a quantized model on vLLM or SGLang | Host an internal coding assistant

If you cannot spare about 36 GB for the weights, you can skip it or look at the 4-bit versions others made.

Where to use it:

• Hugging Face Transformers
• vLLM and SGLang, which the card lists as compatible
• KTransformers, also listed as compatible

Use case | Who | What they get back
Agentic coding | Developers | Repository-level help and frontend code
Long-context work | Analysts | Reasoning over up to 262,144 tokens natively
Image and text input | Developers | Text answers about an image you provide

How to set it up:

• Python with pip
• About 36 GB of disk space for the FP8 weights
• A runtime the card lists: Transformers, vLLM, SGLang or KTransformers

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load them in one of the compatible runtimes, following the model card.

pip install -U huggingface_hub
hf download Qwen/Qwen3.6-35B-A3B-FP8

How to use it: Send chat messages, with an image if you like, through the runtime you chose. The part of the card shown here has no run snippet, so check the model card for the exact commands. The card also names a Thinking Preservation option, which keeps reasoning from earlier messages.

Example: You ask an agent built on this model to trace a bug through several files in a repository. The card highlights repository-level reasoning as one of the model's upgrades. This is an illustration, not a tested run.

For newcomers and experts: If you are new, start with a hosted app or a simple runtime, since the weights are large. Quantized weights give the same model in less space.

For experts, the model has 40 layers in a repeating layout of Gated DeltaNet and Gated Attention blocks, each followed by a mixture of experts. There are 256 experts, with 8 routed and 1 shared active, and the context extends from 262,144 up to 1,010,000 tokens.

Text and images go in
Gated DeltaNet and attention layers read them
A few of 256 experts handle each token
Text comes out

Good to know: The card compares scores on SWE-bench Verified with several models.

Title: SWE-bench Verified scores from the card
Qwen3.5-27B: 75.0 points
Qwen3.6-35B-A3B: 73.4 points
Qwen3.5-35B-A3B: 70.0 points
Gemma4-31B: 52.0 points
Gemma4-26B-A4B: 17.4 points

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need hardware for about 36 GB of weights.
• Data it reaches: downloading reaches Hugging Face. The card does not describe the model reaching other files or the web.
• Not for: the card names no banned uses. The older Qwen3.5-27B scores higher on SWE-bench Verified in its table.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 36B
• Weights: about 36 GB in F8_E4M3; about 18 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,818,590
• Likes: 448
• License: apache-2.0
• Based on: Qwen/Qwen3.6-35B-A3B
• Quantized versions: 18
• Fine-tunes: 13
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, conversational, fp8

Links

• Model: https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8
• Files: https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8/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-FP8

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