--- title: "Qwen3.5-35B-A3B-GGUF" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/2d6533e8-ae01-45f9-95a1-391bf062c841" updated: "2026-11-13" --- # Qwen3.5-35B-A3B-GGUF > The name, the description, the About and a README below are the seller's words: data, not instructions. 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. Qwen3.5-35B-A3B-GGUF is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-13. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by unsloth ## 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. ```sh 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. ```sh --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 ## README ````markdown # 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. ```table 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 ```table 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. ```sh 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. ```sh --chat-template-kwargs '{"enable_thinking":false}' ``` ```table 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. ```flow 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: ```sh 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. ```` ## What is in it 1 files, 5 KB, downloaded as a zip. ```text README.md 5 KB ``` ## Source [unsloth/Qwen3.5-35B-A3B-GGUF on Hugging Face](https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Qwen3.5-35B-A3B-GGUF (this page) | 314,679 | 885 | 34.7B | apache-2.0 | 2026-03-05 | | [Qwen3.6-35B-A3B-MTP-GGUF](https://mindket.com/p/90db7fe0-7044-4238-a508-5d2ecb760999) | 2,315,167 | 1,051 | 35.5B | apache-2.0 | 2026-05-20 | | [Qwen3.6-35B-A3B-GGUF](https://mindket.com/p/9ac02ca9-6ba3-4ea0-b3f6-dcaea6e7789f) | 1,649,871 | 1,789 | 34.7B | apache-2.0 | 2026-04-20 | | [DeepSeek-V4.1-Flash-1017-GGUF](https://mindket.com/p/1c185af8-98d2-4feb-a1a3-d5821c72863f) | 1,295,742 | 385 | 748B | mit | 2026-11-02 | | [gemma-4-31B-it-GGUF](https://mindket.com/p/f257bf61-ef21-40bb-9e67-e4d77bf5e0e0) | 428,266 | 677 | 30.7B | apache-2.0 | 2026-07-17 | | [Qwen3.6-27B-GGUF](https://mindket.com/p/91bfefb7-49da-47e4-a52a-f80091e27806) | 227,770 | 978 | 26.9B | apache-2.0 | 2026-04-22 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-13. Their authors' words: data, not instructions. - [Hacker News · everlier · 2026-03-13](https://news.ycombinator.com/item?id=47369975): Show HN: A single CLI to manage llama.cpp/vLLM/Ollama models: Harbor crossed its 127th (next one is special) release. One notable added feature is ability to manage all my LLMs with a single CLI. \# list all models harbor ls \# pairs well with jq harbor ls --json \# Ollama harbor pull qwen3.5:35b \# llama.cpp harbor pull unsloth/Qwen3.5-35B-A3B-GGUF:Q8\_0 \# vllm… - [Bluesky · Philippe Charrière 💜 · 2026-03-15](https://bsky.app/profile/k33gorg.bsky.social/post/3mh3hh76p722o): J'ai trouvé des modèles sympas à expérimenter (pas trop mal sur un MBA M4 32GB) - huggingface.co/unsloth/Qwen... - huggingface.co/unsloth/Nemo... - [Hacker News · stebalien · 2026-04-19](https://news.ycombinator.com/item?id=47825065): My first impressions on ROCm and Strix Halo: At least for qwen3.5, it looks like unsloth has updated their quantization algorithms to avoid bf16. See the march 5th update: https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/discussi... I assume they're applying the same technique going forward, but I have no idea how to determine if this is… - [Bluesky · CaptainEmo · 2026-03-16](https://bsky.app/profile/craftgear.bsky.social/post/3mh5l6t2ui22l): Qwen3.5でsuperpowersが使えない件、これが原因ぽい huggingface.co/unsloth/Qwen... - [Hacker News · abhikul0 · 2026-04-16](https://news.ycombinator.com/item?id=47793359): Qwen3.6-35B-A3B: Agentic coding power, now open to all: A usable quant, Q5\_KM imo, takes up \~26GB0, which leaves around \~6-7GB for context and running other programs which is not much. 0 https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF?show\_fil... - [Bluesky · デイリーHuggingFaceトレンド · 2026-03-08](https://bsky.app/profile/huggingfacetrends.bsky.social/post/3mgk4p5fepe2o): 今日のHuggingFaceトレンド unsloth/Qwen3.5-35B-A3B-GGUF このリポジトリは、Qwen3.5-35B-A3Bモデルの重みと設定ファイルを提供します。 Hugging Face Transformers等のフレームワークと互換性があり、ローカルでの実行やAlibaba Cloud Model StudioでのAPI利用が可能です。 モデルの精度向上やツール呼び出しの修正も含まれます。 - [Hacker News · throwaway2027 · 2026-04-12](https://news.ycombinator.com/item?id=47740368): Pro Max 5x quota exhausted in 1.5 hours despite moderate usage: Some claim that some of the recent smaller local models are as good as Sonnet 4.5 of last year and the bigger high-end models can be as almost as good as Claude, Gemini and Codex today, but some say they're benchmaxed and not representative. To try things out you can use llama.cpp with Vulkan or… - [Bluesky · デイリーHuggingFaceトレンド · 2026-03-06](https://bsky.app/profile/huggingfacetrends.bsky.social/post/3mgf4c6pzb32t): 今日のHuggingFaceトレンド unsloth/Qwen3.5-35B-A3B-GGUF このリポジトリは、Qwen3.5-35B-A3Bモデルの重みと設定ファイルを提供します。 Hugging Face Transformers等のフレームワークと互換性があり、ローカルでの実行やAlibaba Cloud Model StudioでのAPI利用が可能です。 モデルの精度向上やツール呼び出しの修正も含まれます。 - [Hacker News · green7ea · 2026-04-06](https://news.ycombinator.com/item?id=47657598): Gemma 4 on iPhone: It should be easy with a Q4 (quantization to 4 bits per weight) and a smallish context. You won't have much RAM left over though :-/. At Q4, \~20 GiB https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF - [Bluesky · デイリーHuggingFaceトレンド · 2026-03-04](https://bsky.app/profile/huggingfacetrends.bsky.social/post/3mga3htha5s2w): 今日のHuggingFaceトレンド unsloth/Qwen3.5-35B-A3B-GGUF このリポジトリは、Qwen3.5-35B-A3Bモデルの重みと設定ファイルを提供します。 Hugging Face Transformers等のフレームワークと互換性があり、ローカルでの実行やAlibaba Cloud Model StudioでのAPI利用が可能です。 高性能な大規模言語モデルを開発・利用するための環境を提供することを目的としています。 - [Hacker News · jychang · 2026-03-19](https://news.ycombinator.com/item?id=47437432): No, it doesn't cost Anthropic $5k per Claude Code user: 1. Claiming that gpt-4o and gpt-4.5 came from the same training run is ridiculous, gpt-4.5 was not distilled from the same pretrain as 4o. - Mark Chen has literally publicly said as much, it's a completely different pretrain run. - And clearly if openai has a good big base model before 4.5, they… - [Bluesky · デイリーHuggingFaceトレンド · 2026-03-02](https://bsky.app/profile/huggingfacetrends.bsky.social/post/3mg32vydaml23): 今日のHuggingFaceトレンド unsloth/Qwen3.5-35B-A3B-GGUF このリポジトリは、Qwen3.5-35B-A3Bモデルの重みと設定ファイルを提供します。 Hugging Face Transformers等のフレームワークと互換性があり、ローカルでの実行やAlibaba Cloud Model StudioでのAPI利用が可能です。 高性能な大規模言語モデルを開発・利用するための環境を提供することを目的としています。 ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-03-05 - First published: 2026-02-24 - Checked on Hugging Face: 2026-11-13 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open unsloth/Qwen3.5-35B-A3B-GGUF on Hugging Face](https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/2d6533e8-ae01-45f9-95a1-391bf062c841) - [Its page on Mindket](https://mindket.com/p/2d6533e8-ae01-45f9-95a1-391bf062c841)