Instructions to use ubergarm/Qwen3.5-397B-A17B-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/Qwen3.5-397B-A17B-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/Qwen3.5-397B-A17B-GGUF:Q2_K
# Run inference directly in the terminal:
llama cli -hf ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K
# Run inference directly in the terminal:
llama cli -hf ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K

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/Qwen3.5-397B-A17B-GGUF:Q2_K
# Run inference directly in the terminal:
./llama-cli -hf ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K

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/Qwen3.5-397B-A17B-GGUF:Q2_K
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K

Use Docker

docker model run hf.co/ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K

LM Studio
Jan

vLLM

How to use ubergarm/Qwen3.5-397B-A17B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ubergarm/Qwen3.5-397B-A17B-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/Qwen3.5-397B-A17B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K

Ollama
How to use ubergarm/Qwen3.5-397B-A17B-GGUF with Ollama:
```
ollama run hf.co/ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K
```
Unsloth Desktop

How to use ubergarm/Qwen3.5-397B-A17B-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/Qwen3.5-397B-A17B-GGUF:Q2_K

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/Qwen3.5-397B-A17B-GGUF:Q2_K"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ubergarm/Qwen3.5-397B-A17B-GGUF with Docker Model Runner:
```
docker model run hf.co/ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K
```

Lemonade

How to use ubergarm/Qwen3.5-397B-A17B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ubergarm/Qwen3.5-397B-A17B-GGUF:Q2_K

Run and chat with the model

lemonade run user.Qwen3.5-397B-A17B-GGUF-Q2_K

List all available models

lemonade list

Hermes Agent

How to use ubergarm/Qwen3.5-397B-A17B-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/Qwen3.5-397B-A17B-GGUF:Q2_K

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/Qwen3.5-397B-A17B-GGUF:Q2_K

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ubergarm/Qwen3.5-397B-A17B-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/Qwen3.5-397B-A17B-GGUF:Q2_K

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/Qwen3.5-397B-A17B-GGUF:Q2_K" \
  --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"

Any plans for reviving this model with MTP support?

#14

by phakio - opened May 22

Discussion

phakio

May 22

•

edited May 22

I've been playing with bartowski's quants, and on IQ1_S with MTP full gpu offload, I get nearly 70 t/s TG and 600 t/s PP!

The fun part comes when I tested his IQ4_XS quant on ik_llama.cpp... I get a good 33 t/s TG and 150t/s PP... that's with CPU-MOE offload, leaving lots of room for context on the VRAM. .This, in my opinion, is very usable. I feel as though this model is still relevant today, as artificial analysis places it at 4th place out of all open source models, with reasoning disabled... this shows that it has quite good knowledge density for its size while not relying on reasoning.

I think it's worth revisiting and giving a little life to this guy!

ubergarm

Owner May 24

Heya @phakio

Yes, the whole Qwen3.5 series is pretty amazing. With them teasing Qwen3.7 however I'd probably just wait to see if they drop an updated version and put the effort on that.

There was another thread with folks talking about MTP, but are you able to run one of these models using the MTP tensors from another repo on ik_llama.cpp?

phakio

May 25

I'm hoping they do release another big model with the 3.7 release, but there's a lot of speculation that this model performs too similar to their paid model, and they might just stick to open sourcing the smaller ones from here on out. we'll see soon though!

I saw an interesting comparison chart here https://kaitchup.substack.com/p/lessons-from-gguf-evaluations-ternary , and I'm inclined to heavily agree, that this model is heavily resilient to quantization, as compared to baseline, it only really sees a less than 20% error increase at IQ1... From my testing, even at IQ1 quants the model can do light coding work, accurately reason and solve EVERY medical exam question I come across and feed it, and overall is a really solid model.

I'm not saying it's the best, but I truly do think that with 96GB VRAM, this model, even at IQ1, is a solid choice many may overlook.

Trying to load bartowski's external MTP results in:

llama_model_load: error loading model: check_tensor_dims: tensor 'blk.0.attn_norm.weight' not found
llama_model_load_from_file: failed to load model

phakio changed discussion status to closed May 25

ubergarm

Owner May 26

@phakio

I was running the 122B in 96GB VRAM before all the Qwen3.7-27B MTP craze hit the world haha... I agree the qwen models are pretty resiliant to quantization given the recurrant delta net layers (keeping ssm higher BPW doesn't cost much size) is efficient.

here is an interesting one to try out by a new ik_llama.cpp quantizer @cafkafk with MTP (no imatrix, yet): https://huggingface.co/cafkafk/gemma-4-31B-it-assistant-GGUF-noimatrix

cheers!

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