Instructions to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

LM Studio
Jan

vLLM

How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Ollama
How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with Ollama:
```
ollama run hf.co/unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL
```
Unsloth Desktop

How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.Qwen3-VL-30B-A3B-Thinking-GGUF-UD-Q4_K_XL

List all available models

lemonade list

Hermes Agent

How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF:UD-Q4_K_XL" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Conflicting sampling parameter recommendations?

by ayylmaonade - opened Nov 4, 2025

Discussion

ayylmaonade

Nov 4, 2025

•

edited Nov 4, 2025

Hey! Firstly, thanks for the great work with the UD quants, as always. The Q4_K_XL GGUF is running great at 96K context on my 7900 XTX via llama.cpp w/ Vulkan.

However I've got some questions regarding sampling params, specifically temperature. On your guys' site, the Qwen3-VL guide states to set a temperature of 1.0 for 'Thinking' variants of this model series. But when you click the source, the Qwen3-VL git seems to recommend a temp of 0.6 instead. I wanted to see if I could get some clarification on this if possible?

I've been testing both 0.6 and 1.0. Using temp=1 has caused a few issues for me. I've had one instance of chinese characters showing up, and a few weird cases where the model contains its chain-of-thought outside of the tag. For reference, I'm using the other recommended params;

top_p=0.95
top_k=20
repeat_penalty=1.0

Although, to reduce excessive thinking, I am using presence_penalty=1.5 as was recommended for non-VL Qwen3-Thinking models. I was also wondering if min_p should be set to 0.0 or kept at 0.01? (the llama.cpp default as far as I can tell)

Once again, appreciate all the awesome work you guys do! UD quants are my go-to, and I'd appreciate any advice on this.

Cheers :)

donniel

Dec 22, 2025

Hey! Firstly, thanks for the great work with the UD quants, as always. The Q4_K_XL GGUF is running great at 96K context on my 7900 XTX via llama.cpp w/ Vulkan.

However I've got some questions regarding sampling params, specifically temperature. On your guys' site, the Qwen3-VL guide states to set a temperature of 1.0 for 'Thinking' variants of this model series. But when you click the source, the Qwen3-VL git seems to recommend a temp of 0.6 instead. I wanted to see if I could get some clarification on this if possible?

I've been testing both 0.6 and 1.0. Using temp=1 has caused a few issues for me. I've had one instance of chinese characters showing up, and a few weird cases where the model contains its chain-of-thought outside of the tag. For reference, I'm using the other recommended params;

top_p=0.95

top_k=20

repeat_penalty=1.0

Although, to reduce excessive thinking, I am using presence_penalty=1.5 as was recommended for non-VL Qwen3-Thinking models. I was also wondering if min_p should be set to 0.0 or kept at 0.01? (the llama.cpp default as far as I can tell)

Once again, appreciate all the awesome work you guys do! UD quants are my go-to, and I'd appreciate any advice on this.

Cheers :)

@ayylmaonade I'm running into the same issue! I'm wondering if you stuck with 0.6, and if that worked better for you? What about presence_penalty and min_p?

ayylmaonade

Dec 24, 2025

Hey! Firstly, thanks for the great work with the UD quants, as always. The Q4_K_XL GGUF is running great at 96K context on my 7900 XTX via llama.cpp w/ Vulkan.

However I've got some questions regarding sampling params, specifically temperature. On your guys' site, the Qwen3-VL guide states to set a temperature of 1.0 for 'Thinking' variants of this model series. But when you click the source, the Qwen3-VL git seems to recommend a temp of 0.6 instead. I wanted to see if I could get some clarification on this if possible?

I've been testing both 0.6 and 1.0. Using temp=1 has caused a few issues for me. I've had one instance of chinese characters showing up, and a few weird cases where the model contains its chain-of-thought outside of the tag. For reference, I'm using the other recommended params;

top_p=0.95

top_k=20

repeat_penalty=1.0

Although, to reduce excessive thinking, I am using presence_penalty=1.5 as was recommended for non-VL Qwen3-Thinking models. I was also wondering if min_p should be set to 0.0 or kept at 0.01? (the llama.cpp default as far as I can tell)

Once again, appreciate all the awesome work you guys do! UD quants are my go-to, and I'd appreciate any advice on this.

Cheers :)

@ayylmaonade I'm running into the same issue! I'm wondering if you stuck with 0.6, and if that worked better for you? What about presence_penalty and min_p?

After a few days of tinkering with a whole assortment of settings, I finally landed on a decent combination. It’s still a notch below the older 2507 checkpoint when it comes to pure‑text tasks though. I settled on a temperature of 1.0, a min_p of 0.0, and a presence_penalty of 1.5. Even though a few repos suggest a penalty of 0.0, I saw no drop in quality and over‑thinking got noticeably cut down. All the other parameters stayed the same: top_k 20, top_p 0.95, repeat_penalty 1.0.

I also found a handy benchmark that directly compares VL models against their non‑VL counterparts in text‑based scenarios. https://dubesor.de/benchtable - have a look here if you want, just search "Qwen3" - at the very least it confirms I'm not insane regarding the reduction in text-performance, lol.

donniel

Feb 26

Thanks! Missed this notification, but I'm on GLM 4.7 Flash and currently evaluating Qwen3.5 now 😁

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment