Instructions to use ubergarm/Qwen3.8-27B-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 ubergarm/Qwen3.8-27B-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 ubergarm/Qwen3.8-27B-GGUF
# Run inference directly in the terminal:
llama cli -hf ubergarm/Qwen3.8-27B-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ubergarm/Qwen3.8-27B-GGUF
# Run inference directly in the terminal:
llama cli -hf ubergarm/Qwen3.8-27B-GGUF

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 ubergarm/Qwen3.8-27B-GGUF
# Run inference directly in the terminal:
./llama-cli -hf ubergarm/Qwen3.8-27B-GGUF

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 ubergarm/Qwen3.8-27B-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ubergarm/Qwen3.8-27B-GGUF

Use Docker

docker model run hf.co/ubergarm/Qwen3.8-27B-GGUF

LM Studio
Jan

vLLM

How to use ubergarm/Qwen3.8-27B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ubergarm/Qwen3.8-27B-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": "ubergarm/Qwen3.8-27B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ubergarm/Qwen3.8-27B-GGUF

Ollama
How to use ubergarm/Qwen3.8-27B-GGUF with Ollama:
```
ollama run hf.co/ubergarm/Qwen3.8-27B-GGUF
```
Unsloth Desktop

How to use ubergarm/Qwen3.8-27B-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ubergarm/Qwen3.8-27B-GGUF

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": "ubergarm/Qwen3.8-27B-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ubergarm/Qwen3.8-27B-GGUF with Docker Model Runner:
```
docker model run hf.co/ubergarm/Qwen3.8-27B-GGUF
```

Lemonade

How to use ubergarm/Qwen3.8-27B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ubergarm/Qwen3.8-27B-GGUF

Run and chat with the model

lemonade run user.Qwen3.8-27B-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use ubergarm/Qwen3.8-27B-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 ubergarm/Qwen3.8-27B-GGUF

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 ubergarm/Qwen3.8-27B-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ubergarm/Qwen3.8-27B-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ubergarm/Qwen3.8-27B-GGUF

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 "ubergarm/Qwen3.8-27B-GGUF" \
  --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"

Time taken to create this quant

by igormp - opened Aug 18

Discussion

igormp

Aug 18

Hello! First of all, thanks for providing those IK quants! Those quants you make are often my favorite ones in terms of quality and performance for a higher-end-ish but still consumer device 😃

As for my actual question: how long does it usually take you to create those quantizations on your personal setup? IIRC you did this one without using L1T's hardware, right?
I'm trying to get an estimate of the time I'd need so I could give it a go at creating my own versions using your recipes.

ubergarm

Owner Aug 18

As for my actual question: how long does it usually take you to create those quantizations on your personal setup?

This one took only a couple hours. You can look in the logs directory for full commands and expected outputs, and adjust the recipe to suit your personal needs. Feel free to use my imatrix file to save some time, as that is typically the most expensive / slowest step.

download original bf16 safetensors
mainline llama.cpp convert_hf_to_gguf.py
collect imatrix of bf16 GGUF using ik_llama.cpp
quantize desired recipe
run --validate-quants, perplexity test, and vibe check with pi harness
upload to hf

IIRC you did this one without using L1T's hardware, right?

Yep, did this one with my home gaming rig. This is my build (I only bought a single 3090 though, and still regret not picking up the second one for $500 when I had the chance lol) https://pcpartpicker.com/b/tMsXsY

If you have a fast PCIe Gen 5 NVMe drive, and fast DDR5-6400 MT/s RAM, and a newer AMD CPU with avx512_vnni, it is plenty good for these "smaller" LLMs.

KeinNiemand

Aug 27

•

edited Aug 27

I'm currently making an IQ_K quant set for dsv4 flash 0731 on my gaming rig (7800x3d, ddr5-6000, 4.0 nvme ssd) and it's taking me
like 2-6h per quant (i run these overnight and don't measure how long they take so don't really know)+
like 2-3h downloading the saftensors with gigabit ethernet+ 2
h or so running convert_hf_to_gguf.py
not sure how it scales with parameter count and if it's linear or not but of course a 27B would be way faster then a 304B

the real thing that has expanded my dsv4 flash 0731 run from like one 12-16h overnight+work time run to multiple weeks of working on it now is that somebody provided a kld file for the perfect version which means i can actually measure my quants kld against a reference which in turn means i could see that my intimal attempt was worse then atomic chats quants that where at the same size => which simply doesn't do the whole point of IQ_K quants is to be better and more efficient then mainline quants at the same size so had to redo the recipe and rerun the entire quant set several time to get it to work.
I'm finally at a point I'm happy with now all I have left to do is make a readme and spend another many hours uploading the whole thing.

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