Instructions to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ 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 w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16
# Run inference directly in the terminal:
llama cli -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16
# Run inference directly in the terminal:
llama cli -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

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 w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16
# Run inference directly in the terminal:
./llama-cli -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

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 w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

Use Docker

docker model run hf.co/w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

LM Studio
Jan

vLLM

How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "w-ahmad/Qwen3.5-9B-GGUF-MoQ"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "w-ahmad/Qwen3.5-9B-GGUF-MoQ",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

Ollama
How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with Ollama:
```
ollama run hf.co/w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16
```

Unsloth Studio

How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for w-ahmad/Qwen3.5-9B-GGUF-MoQ to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for w-ahmad/Qwen3.5-9B-GGUF-MoQ to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for w-ahmad/Qwen3.5-9B-GGUF-MoQ to start chatting

How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

OpenClaw new

How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

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 "w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16" \
  --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"

Docker Model Runner
How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with Docker Model Runner:
```
docker model run hf.co/w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16
```

Lemonade

How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

Run and chat with the model

lemonade run user.Qwen3.5-9B-GGUF-MoQ-BF16

List all available models

lemonade list

Hermes Agent

How to use w-ahmad/Qwen3.5-9B-GGUF-MoQ with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

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 w-ahmad/Qwen3.5-9B-GGUF-MoQ:BF16

Run Hermes

hermes

Atomic Chat

Qwen3.5-9B-GGUF-MoQ / README.md

w-ahmad

Update README.md

93149d6 verified 3 months ago

7.11 kB

language:
  - en
library_name: gguf
tags:
  - MoQ
  - mixture-of-quants
  - GGUF
  - QWEN
  - quantization
base_model:
  - Qwen/Qwen3.5-9B
license: mit
pipeline_tag: text-generation

🚀 MoQ: Mixture of Quants

MoQ (Mixture of Quants) is a smart way to shrink AI models without losing their "brainpower." Unlike old methods that treat every part of the model the same, MoQ identifies the most important parts and keeps them high-quality, while heavily compressing the rest to save space.**Stop settling for uniform bitrates. Standard quantization is a relic of the past, treating vital cognitive weights the same as redundant noise. MoQ (Mixture of Quants) is a surgical evolution in model compression. By deploying an Empirical Per-Tensor Analysis, MoQ identifies the "High-Intelligence" tensors that drive reasoning and shields them with high-bit precision, while crushing redundant weights into extreme efficiency.

The result? A model that punches significantly above its weight class.

📥 Available Quants

Folder Link	BPW	Total Size
📂 BF-16	16	~16.69 GB
📂 FP-16	16	~16.69 GB
📂 MoQ-2.10	2.10	~2.23 GB
📂 MoQ-2.55	2.55	~2.68 GB
📂 MoQ-2.80	2.80	~2.90 GB
📂 MoQ-2.85	2.85	~3.00 GB
📂 MoQ-2.90	2.90	~3.03 GB
📂 MoQ-3.10	3.10	~3.24 GB
📂 MoQ-3.25	3.25	~3.41 GB
📂 MoQ-3.30	3.30	~3.44 GB
📂 MoQ-3.65	3.65	~3.79 GB
📂 MoQ-3.75	3.75	~3.92 GB
📂 MoQ-4.00	4.00	~4.20 GB
📂 MoQ-4.25	4.25	~4.43 GB
📂 MoQ-4.65	4.65	~4.87 GB
📂 MoQ-4.85	4.85	~5.05 GB
📂 MoQ-5.00	5.00	~5.24 GB
📂 MoQ-5.65	5.65	~5.91 GB
📂 MoQ-6.35	6.35	~6.61 GB
📂 MoQ-7.80	7.80	~8.17 GB

📊 MoQ vs. Unsloth: Performance Comparison

The following benchmarks compare MoQ 4.85 against Unsloth Dynamic Quants. Note: Lower KLD (Kullback–Leibler Divergence) indicates higher fidelity to the original model.

📉 Key Divergence Metrics (Lower is Better)

Mean KLD

Average divergence across all layers. MoQ maintains a significantly lower average error profile.

Maximum KLD

The "worst-case" divergence point. MoQ 4.84 effectively eliminates the extreme divergence spikes seen in standard dynamic quants.

RMS Δp

Root Mean Square change in probabilities. This measures the stability of the model's confidence.

📈 Precision Percentiles

These graphs demonstrate MoQ's ability to maintain stability even within the most sensitive portions of the architecture.

95.0% Percentile KLD

99.0% Percentile KLD

99.9% Percentile KLD

🎯 Token Match (Same Top-P)

This metric tracks how often the quantized model chooses the exact same top token as the original high-precision model.

🧠 The MoQ Edge

MoQ optimizes the architecture for the Pareto frontier of memory and performance.

Dynamic Bitrate Allocation: No more "one-size-fits-all." MoQ assigns precision where it actually matters.
Cognitive Preservation: Massive VRAM savings with near-zero degradation in logic and coherence.
Next-Gen Efficiency: Fits "Large" model intelligence into "Small" model hardware.

📍 Weight Sensitivity Heatmap

Lighter regions represent mission-critical tensors preserved at higher precision.

📈 Importance Distribution

The histogram shows the importance scores used to mathematically determine the optimal quant for each tensor.

Follow me on Linkedin linkedin.com/in/waleed-ahmad-8a3166403 If MoQ does not perform well, email me : waleedahmad.1a10@gmail.com

🛠 Usage & Deployment

./llama-cli -m Qwen3.5-9B-MoQ-4.85.gguf -p "The future of efficient AI is..."