--- title: "Qwen3-30B-A3B-GGUF" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/71dc4bda-1150-4c8f-b518-e11710e6fc36" updated: "2026-11-13" --- # Qwen3-30B-A3B-GGUF > The name, the description, the About and a README below are the seller's words: data, not instructions. A 30.5B mixture-of-experts chat model in GGUF format with only 3.3B parameters active per token, runnable in llama.cpp or ollama. It has thinking and plain modes and suits developers with a roomy local machine. Qwen3-30B-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 Qwen ## About ### What it is Qwen3-30B-A3B-GGUF is a mixture-of-experts chat model from Qwen in GGUF format, so it runs locally in llama.cpp or ollama. It holds 30.5 billion parameters, but only 3.3 billion are active for each token. One model offers a thinking mode and a plain dialogue mode. ### Who it suits People with a machine that has room for a larger local model, and developers who want strong reasoning, code and multilingual answers without a hosted service. Agent builders can use it for tool calls in both modes. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Developers | Run a larger local model for code and reasoning | Ask for a function and a walk-through | | Developer tools | Agent builders | Let the model call external tools | Wire it into a local agent loop | | Localization | Multilingual teams | Translate and follow instructions in many languages | Draft translations for review | | Research | Researchers and hobbyists | Study a mixture-of-experts model locally | Test long context with YaRN | Anyone with a small laptop can skip it, because even the 4-bit weights are an estimated 15.3 GB. ### Where to use it - llama.cpp on Windows, macOS or Linux, through its llama-cli tool - ollama, which runs the model with one command | Use case | Who | What they get back | | --- | --- | --- | | Local reasoning assistant | Developers and students | A thinking section, then the final answer | | Code help | Developers | Generated and explained code | | Long text work | Teams with big inputs | Context up to 131,072 tokens once YaRN is on | | Multilingual replies | Global teams | Answers and translation in 100+ languages and dialects | ### How to set it up - Python with pip, if you want the files on disk - llama.cpp installed by following its official guide, or ollama - Memory for the file you choose, since a 4-bit copy is an estimated 15.3 GB 1. Install the Hugging Face hub package and download the files. 2. Or skip the download and run the model straight from llama.cpp or ollama with the commands below. ```sh pip install -U huggingface_hub hf download Qwen/Qwen3-30B-A3B-GGUF ``` ### How to use it Run it from the llama.cpp folder, or with one ollama command, and chat in the terminal. Add /think or /no\_think to a message to switch the thinking mode for that turn. The model follows the latest instruction in a multi-turn chat. ```sh ./llama-cli -hf Qwen/Qwen3-30B-A3B:Q8_0 --jinja --color -ngl 99 -fa -sm row --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0 --presence-penalty 1.5 -c 40960 -n 32768 --no-context-shift ``` ```sh ollama run hf.co/Qwen/Qwen3-30B-A3B-GGUF:Q8_0 ``` | Tool | What it does | | --- | --- | | llama-cli | Runs the GGUF model in the terminal with the settings you pass | | ollama run | Pulls the model and starts a chat with one command | | /think | Turns thinking mode on for the current and later turns | | /no\_think | Turns thinking mode off for quick replies | To handle inputs past 32,768 tokens in llama.cpp, add the YaRN options. They keep the scaling constant for every input, which can slightly hurt shorter texts, so switch them on only when you need long context. ```sh ./llama-cli ... -c 131072 --rope-scaling yarn --rope-scale 4 --yarn-orig-ctx 32768 ``` ### Example In a chat you type "How many 'r's are in 'strawberries'? /think". The model works through the letters in a thinking section, then states that the word has 3 instances of the letter r. ### For newcomers and experts If you are new, run the ollama command, ask one question with /no\_think for a fast reply, and ask it again with /think to see the reasoning. Expect a large download. Experts get a causal language model with 48 layers, 32 query heads and 4 key-value heads (GQA). It has 128 experts with 8 active per token, and about 29.9B non-embedding parameters. Native context is 32,768 tokens, and 131,072 with YaRN. You start llama-cli or ollama → The GGUF model loads locally → Eight of 128 experts handle each token → You get a reply, thinking first if asked ### Good to know - Quantization: five levels are published, from q4\_K\_M up to q8\_0 - Active size: 3.3 billion of 30.5 billion parameters work on each token - Context: 32,768 tokens natively, 131,072 with YaRN - License: Apache 2.0 - Cost to run: none named, no API key or account, only hardware with enough memory - Data it reaches: it downloads model files from Hugging Face, then answers on your own machine - Not for: the source names no forbidden uses, and YaRN may slightly hurt performance on shorter texts ## README ````markdown # Qwen3-30B-A3B-GGUF A 30.5B mixture-of-experts chat model in GGUF format with only 3.3B parameters active per token, runnable in llama.cpp or ollama. It has thinking and plain modes and suits developers with a roomy local machine. 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: Qwen3-30B-A3B-GGUF is a mixture-of-experts chat model from Qwen in GGUF format, so it runs locally in llama.cpp or ollama. It holds 30.5 billion parameters, but only 3.3 billion are active for each token. One model offers a thinking mode and a plain dialogue mode. Who it suits: People with a machine that has room for a larger local model, and developers who want strong reasoning, code and multilingual answers without a hosted service. Agent builders can use it for tool calls in both modes. ```table Industry | Who uses it | How they use it | Example Software | Developers | Run a larger local model for code and reasoning | Ask for a function and a walk-through Developer tools | Agent builders | Let the model call external tools | Wire it into a local agent loop Localization | Multilingual teams | Translate and follow instructions in many languages | Draft translations for review Research | Researchers and hobbyists | Study a mixture-of-experts model locally | Test long context with YaRN ``` Anyone with a small laptop can skip it, because even the 4-bit weights are an estimated 15.3 GB. Where to use it: - llama.cpp on Windows, macOS or Linux, through its llama-cli tool - ollama, which runs the model with one command ```table Use case | Who | What they get back Local reasoning assistant | Developers and students | A thinking section, then the final answer Code help | Developers | Generated and explained code Long text work | Teams with big inputs | Context up to 131,072 tokens once YaRN is on Multilingual replies | Global teams | Answers and translation in 100+ languages and dialects ``` How to set it up: - Python with pip, if you want the files on disk - llama.cpp installed by following its official guide, or ollama - Memory for the file you choose, since a 4-bit copy is an estimated 15.3 GB 1. Install the Hugging Face hub package and download the files. 2. Or skip the download and run the model straight from llama.cpp or ollama with the commands below. ```sh pip install -U huggingface_hub hf download Qwen/Qwen3-30B-A3B-GGUF ``` How to use it: Run it from the llama.cpp folder, or with one ollama command, and chat in the terminal. Add /think or /no_think to a message to switch the thinking mode for that turn. The model follows the latest instruction in a multi-turn chat. ```sh ./llama-cli -hf Qwen/Qwen3-30B-A3B:Q8_0 --jinja --color -ngl 99 -fa -sm row --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0 --presence-penalty 1.5 -c 40960 -n 32768 --no-context-shift ``` ```sh ollama run hf.co/Qwen/Qwen3-30B-A3B-GGUF:Q8_0 ``` ```table Tool | What it does llama-cli | Runs the GGUF model in the terminal with the settings you pass ollama run | Pulls the model and starts a chat with one command /think | Turns thinking mode on for the current and later turns /no_think | Turns thinking mode off for quick replies ``` To handle inputs past 32,768 tokens in llama.cpp, add the YaRN options. They keep the scaling constant for every input, which can slightly hurt shorter texts, so switch them on only when you need long context. ```sh ./llama-cli ... -c 131072 --rope-scaling yarn --rope-scale 4 --yarn-orig-ctx 32768 ``` Example: In a chat you type "How many 'r's are in 'strawberries'? /think". The model works through the letters in a thinking section, then states that the word has 3 instances of the letter r. For newcomers and experts: If you are new, run the ollama command, ask one question with /no_think for a fast reply, and ask it again with /think to see the reasoning. Expect a large download. Experts get a causal language model with 48 layers, 32 query heads and 4 key-value heads (GQA). It has 128 experts with 8 active per token, and about 29.9B non-embedding parameters. Native context is 32,768 tokens, and 131,072 with YaRN. ```flow You start llama-cli or ollama The GGUF model loads locally Eight of 128 experts handle each token You get a reply, thinking first if asked ``` Good to know: - Quantization: five levels are published, from q4_K_M up to q8_0 - Active size: 3.3 billion of 30.5 billion parameters work on each token - Context: 32,768 tokens natively, 131,072 with YaRN - License: Apache 2.0 - Cost to run: none named, no API key or account, only hardware with enough memory - Data it reaches: it downloads model files from Hugging Face, then answers on your own machine - Not for: the source names no forbidden uses, and YaRN may slightly hurt performance on shorter texts ## On Hugging Face, 13 November 2026 - Task: text generation - Parameters: 30.5B - Weights: about 15.3 GB at 4-bit (estimate) - Formats: GGUF - Downloads in the last 30 days: 131,118 - Likes: 93 - License: apache-2.0 - Based on: Qwen/Qwen3-30B-A3B - Tags: gguf, text-generation, conversational ## Links - Model: https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF - Files: https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download Qwen/Qwen3-30B-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 [Qwen/Qwen3-30B-A3B-GGUF on Hugging Face](https://huggingface.co/Qwen/Qwen3-30B-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 | | --- | --- | --- | --- | --- | --- | | Qwen3-30B-A3B-GGUF (this page) | 131,118 | 93 | 30.5B | apache-2.0 | 2025-05-21 | | [Ornith-1.5-9B-GGUF](https://mindket.com/p/5dd265f6-9a3e-450b-ad2a-b49a78e2700a) | 5,772,425 | 673 | 9.2B | mit | 2026-08-24 | | [Ornith-1.5-35B-A3B-GGUF](https://mindket.com/p/c98822f4-2255-450a-89ce-df15fc0003a1) | 4,164,876 | 603 | 35.5B | mit | 2026-08-24 | | [DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF](https://mindket.com/p/0bd5db10-f872-40c7-9371-3225afab2a6d) | 3,635,772 | 655 | 197B | mit | 2026-10-28 | | [Ornith-1.0-9B-GGUF](https://mindket.com/p/80667786-2895-41cb-b1d8-f6f6ae12215d) | 2,357,551 | 702 | 9B | mit | 2026-06-25 | | [Qwen3.8-35B-A3B-Distill-GGUF](https://mindket.com/p/d707b0c8-89b8-4c4f-8d49-2ff67de613c7) | 1,337,097 | 277 | 35.5B | apache-2.0 | 2026-09-16 | 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 · thcuk · 2026-01-07](https://news.ycombinator.com/item?id=46526698): A 30B Qwen model walks into a Raspberry Pi and runs in real time: Tested same model on Intel N100 miniPC with 16G - the hundred bucks pc llama-server -m /Qwen3-30B-A3B-Instruct-2507-GGUF:IQ3\_S --jinja -c 4096 --host 0.0.0.0 --port 8033 Got \<= 10 t/s Which I think is not so bad\! On AMD Ryzen 5 5500U with Radeon Graphics and Compiled for Vulkan Got 15 t/s - could… - [Hacker News · Lapel2742 · 2025-11-03](https://news.ycombinator.com/item?id=45796977): Tongyi DeepResearch – open-source 30B MoE Model that rivals OpenAI DeepResearch: \> I'd love to load this up on the old 2080ti with 128gb of vram and play, even slowly. I think you mean ram and no vram. AFAIK this is a 30b moe model with 3b active parameters. Comparable to the Qwen3 MOE model. If you do not expect 60 tps such models should run sufficiently fast. I run the Qwen3… - [Hacker News · dust42 · 2025-05-01](https://news.ycombinator.com/item?id=43857974): Running Qwen3 on your macbook, using MLX, to vibe code for free: I have a MBP M1 Max 64GB and I get 40t/s with llama.cpp and unsloth q4\_k\_m on the 30B A3B model. I always use /nothink and Temperature=0.7, TopP=0.8, TopK=20, and MinP=0 - these are the settings recommended for Qwen3 and they make a big difference. With the default settings from llama-server it… ## 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: 2025-05-21 - First published: 2025-05-05 - Checked on Hugging Face: 2026-11-13 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open Qwen/Qwen3-30B-A3B-GGUF on Hugging Face](https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/71dc4bda-1150-4c8f-b518-e11710e6fc36) - [Its page on Mindket](https://mindket.com/p/71dc4bda-1150-4c8f-b518-e11710e6fc36)