Qwen3.5-35B-A3B-GGUF

Qwen3.5-35B-A3B-GGUF

Reads text and images, writes code and calls tools on your own machine, as Unsloth's refreshed GGUF files for the Qwen3.5 35B-A3B vision-language model. For developers who run local agents.

Made by unsloth 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 Unsloth's GGUF copy of Qwen3.5-35B-A3B, a vision-language model from the Qwen team. The record lists about 34.7B parameters, and the A3B in the name points to a sparse mixture-of-experts design with roughly 3B active. Qwen3.5 combines Gated Delta Networks with sparse experts for fast inference. Unsloth rebuilt these files with new imatrix data and an improved quantization algorithm.

Who it suits: It suits developers who want a local model that reads images, writes code and calls tools, and teams that already use Claude Code or Codex.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local assistant for coding, images and tool calls | Use it behind Claude Code or Codex
Research | ML researchers | Compare quantization variants of one model | Test Q4_K_M against Q8_0
Self-hosting | Teams with large-memory machines | Keep prompts on their own hardware | Serve the 4-bit files in house

Anyone who prefers a managed service can skip it, since the card points to a hosted Qwen3.5-Flash on Alibaba Cloud.

Where to use it:

• llama.cpp and other tools that read GGUF files
• Claude Code and Codex, where the card says tool calling and coding improved
• Unsloth's free Colab notebooks, for fine-tuning and reinforcement learning
• Windows, macOS or Linux, where the files download with Python

Use case | Who | What they get back
Coding with tool calls | Developers | Code and function calls from a local model
Questions about images | Everyday users | A written answer about a picture
Fine-tune or train with RL | Researchers | A base to adapt with Unsloth

How to set it up:

• Python with pip, for the download
• A runtime that can load GGUF files
• Memory for about 17.3 GB at 4-bit, as an estimate

1. Install the Hugging Face hub client.
2. Download the repository's files, or only the variant you want.
3. Load a file in your runtime and follow Unsloth's guide for the right settings.

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

How to use it: Load a quantization and chat with it. To switch thinking off, pass the template option from the card when you start the server.

--chat-template-kwargs '{"enable_thinking":false}'

Tool | What it does
--chat-template-kwargs | Passes template options to the server, here to turn thinking off
Q8_0, Q4_K_M and BF16 | Variants uploaded in the final refresh
Colab notebooks | Free notebooks for fine-tuning and reinforcement learning with Unsloth

Example: You start the model with thinking switched off and ask it to explain a function you paste. It answers directly, without a reasoning phase first.

For newcomers and experts: Load a 4-bit file, ask a plain question and see how it feels before you add thinking or tools.

Qwen3.5 trains on text and image tokens together from the start, and pairs Gated Delta Networks with sparse experts. For this size, Unsloth says its last refresh mainly lowered the maximum KLD, a measure of how far the quantized output drifts from the original. It also removed MXFP4 layers from three of the quantizations.

Text and images go in
A hybrid of delta networks and sparse experts reads them
Tool calls and code come out as text
You switch thinking on or off

Good to know:

• Refresh: all files use new imatrix data and an improved quantization algorithm
• Size: the record lists 34.7B parameters, with about 17.3 GB at 4-bit as an estimate
• Hosted option: the card mentions Qwen3.5-Flash, a hosted version with a 1M default context
• License: Apache 2.0
• Cost to run: none named, and the hosted option is separate
• Data it reaches: the text and images you give it, with no web or account access named for the local files
• Not for: nothing named in the card

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 34.7B
• Weights: about 17.3 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 314,679
• Likes: 885
• License: apache-2.0
• Based on: Qwen/Qwen3.5-35B-A3B
• Quantized versions: 1
• Fine-tunes: 1
• Tags: gguf, qwen3_5_moe, unsloth, image-text-to-text, imatrix, conversational

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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