Instructions to use Majipa/Qwen3-0.6B-GGUF-Quantized with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Majipa/Qwen3-0.6B-GGUF-Quantized with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Majipa/Qwen3-0.6B-GGUF-Quantized",
	filename="Qwen3-0.6B-IQ3_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Majipa/Qwen3-0.6B-GGUF-Quantized 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 Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

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 Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

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 Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

Use Docker

docker model run hf.co/Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

LM Studio
Jan

vLLM

How to use Majipa/Qwen3-0.6B-GGUF-Quantized with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Majipa/Qwen3-0.6B-GGUF-Quantized"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Majipa/Qwen3-0.6B-GGUF-Quantized",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

Ollama
How to use Majipa/Qwen3-0.6B-GGUF-Quantized with Ollama:
```
ollama run hf.co/Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M
```

Unsloth Studio

How to use Majipa/Qwen3-0.6B-GGUF-Quantized 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 Majipa/Qwen3-0.6B-GGUF-Quantized 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 Majipa/Qwen3-0.6B-GGUF-Quantized to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Majipa/Qwen3-0.6B-GGUF-Quantized to start chatting

How to use Majipa/Qwen3-0.6B-GGUF-Quantized with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

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": "Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use Majipa/Qwen3-0.6B-GGUF-Quantized with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

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 Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use Majipa/Qwen3-0.6B-GGUF-Quantized with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

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 "Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M" \
  --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 Majipa/Qwen3-0.6B-GGUF-Quantized with Docker Model Runner:
```
docker model run hf.co/Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M
```

Lemonade

How to use Majipa/Qwen3-0.6B-GGUF-Quantized with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Majipa/Qwen3-0.6B-GGUF-Quantized:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3-0.6B-GGUF-Quantized-Q4_K_M

List all available models

lemonade list

Qwen3-0.6B — GGUF Quantizations

GGUF quantizations of Qwen/Qwen3-0.6B-Base, built with llama.cpp for local, CPU-friendly inference. This repo ships 21 quantization formats (from TQ1_0 up to Q8_0) so you can trade size, speed and quality to fit your hardware.

Original model: https://huggingface.co/Qwen/Qwen3-0.6B-Base
Format: GGUF (works with llama.cpp, llama-cpp-python, Ollama, LM Studio, Jan, …)
Reproducible pipeline: https://github.com/majipa007/LLM-Quantization-Lab

These are independently quantized derivatives of the source model. They are base (non-instruct) weights, so they complete text rather than follow chat instructions.

Model details


Base model	`Qwen/Qwen3-0.6B-Base`
Architecture	`Qwen3ForCausalLM` (`qwen3`)
Parameters	~0.6B (hidden size 1024, 28 layers, 16 heads / 8 KV heads)
Vocabulary	151,936
Context length	32,768
Tensor dtype (source)	bfloat16
License	Apache-2.0

Conversion & quantization process

Download the source model from the Hugging Face Hub.
Convert it to a full-precision BF16 GGUF with llama.cpp's convert_hf_to_gguf.py.
Quantize that BF16 baseline into each target format with llama-quantize.

# 1) BF16 baseline
python llama.cpp/convert_hf_to_gguf.py \
  models/hf/Qwen3-0.6B \
  --outfile Qwen3-0.6B-BF16.gguf \
  --outtype bf16

# 2) One quantization (repeat per format)
./llama.cpp/build/bin/llama-quantize \
  Qwen3-0.6B-BF16.gguf \
  Qwen3-0.6B-Q4_K_M.gguf \
  Q4_K_M

Tooling: llama.cpp @ 91c631b21d6e5d09e9c6659efdf6baeef5a44ddb (branch master)
Importance matrix (imatrix): not used — these are plain quantizations.

Available quantizations & file sizes

Sizes are the actual on-disk bytes; “% of BF16” is the size relative to the BF16 baseline (1439.4 MiB). Smaller = more compression = lower quality.

File	Bits (approx)	Size	% of BF16	Notes
`Qwen3-0.6B-Q8_0.gguf`	8.5	767.5 MiB	53.3%	Near-lossless; largest quant
`Qwen3-0.6B-Q6_K.gguf`	6.6	593.9 MiB	41.3%	Very high quality
`Qwen3-0.6B-Q5_1.gguf`	6.0	553.9 MiB	38.5%	Legacy 5-bit
`Qwen3-0.6B-Q5_K_M.gguf`	5.7	525.8 MiB	36.5%	High quality
`Qwen3-0.6B-Q5_K_S.gguf`	5.5	518.4 MiB	36.0%	High quality, slightly smaller
`Qwen3-0.6B-Q5_0.gguf`	5.5	518.4 MiB	36.0%	Legacy 5-bit
`Qwen3-0.6B-Q4_1.gguf`	5.0	482.9 MiB	33.5%	Legacy 4-bit
`Qwen3-0.6B-Q4_K_M.gguf`	4.8	461.8 MiB	32.1%	Recommended default — best size/quality balance
`Qwen3-0.6B-Q4_K_S.gguf`	4.6	449.0 MiB	31.2%	Good balance, smaller
`Qwen3-0.6B-IQ4_NL.gguf`	4.5	448.5 MiB	31.2%	I-quant, non-linear
`Qwen3-0.6B-Q4_0.gguf`	4.5	447.4 MiB	31.1%	Legacy 4-bit
`Qwen3-0.6B-IQ4_XS.gguf`	4.3	431.0 MiB	29.9%	I-quant, good quality-per-byte
`Qwen3-0.6B-Q3_K_L.gguf`	3.9	415.2 MiB	28.8%	3-bit, larger variant
`Qwen3-0.6B-Q3_K_M.gguf`	3.7	394.8 MiB	27.4%	3-bit, medium
`Qwen3-0.6B-IQ3_M.gguf`	3.5	384.2 MiB	26.7%	I-quant 3-bit
`Qwen3-0.6B-IQ3_S.gguf`	3.4	371.9 MiB	25.8%	I-quant 3-bit, smaller
`Qwen3-0.6B-Q3_K_S.gguf`	3.4	371.9 MiB	25.8%	3-bit, small
`Qwen3-0.6B-Q2_K.gguf`	2.6	331.2 MiB	23.0%	2-bit; noticeable quality loss
`Qwen3-0.6B-TQ2_0.gguf`	2.1	319.4 MiB	22.2%	Ternary; experimental
`Qwen3-0.6B-TQ1_0.gguf`	1.7	299.7 MiB	20.8%	Ternary; experimental
`Qwen3-0.6B-Q1_0.gguf`	1.0	207.6 MiB	14.4%	1-bit; extreme compression, large quality loss

Bit-widths are approximate effective bits-per-weight for the format family.

Recommended files

For a 0.6B model, quality degrades quickly at very low bit-widths, so prefer higher bits when you can afford the size:

Best all-round: Q4_K_M — the standard sweet spot.
Want more quality: Q5_K_M or Q6_K.
Maximum fidelity: Q8_0 (effectively lossless vs BF16 for most uses).
Tight on memory: IQ4_XS or Q3_K_M (accept some quality loss).
Experimental / research only: Q2_K, TQ2_0, TQ1_0, Q1_0 — expect significant degradation on a model this small; not recommended for real use.

Speed benchmarks

Pending. Throughput (prompt-processing and text-generation tokens/s) will be measured with llama-bench and added here. Methodology: llama-bench -m <file> -p 512 -n 128 -r 5 on the hardware below, CPU-only (-ngl 0).

File	Prompt (t/s)	Generation (t/s)
to be added

Quality evaluation

Pending. Perplexity on the WikiText-2 raw test set (via llama-perplexity, context 4096) will be reported per quantization, alongside the BF16 baseline, so quality loss can be compared directly. Note perplexity only compares models sharing the same tokenizer and is not a complete measure of practical quality.

File	Perplexity (WikiText-2)	Δ vs BF16
to be added

Hardware specifications (build & reference)

Quantization and (upcoming) benchmarks were produced on:


CPU	13th Gen Intel Core i7-1355U (12 threads: 6 cores × 2)
Memory	16 GB
GPU	none (CPU-only)
OS	Ubuntu 24.04.4 LTS on WSL2
Compiler	gcc/g++ 13.3.0
CMake	3.28.3
Python	3.12

Usage

Replace majipa007/Qwen3-0.6B-GGUF with this repo's id if different.

Download a single file

pip install -U "huggingface_hub[cli]"
hf download majipa007/Qwen3-0.6B-GGUF Qwen3-0.6B-Q4_K_M.gguf --local-dir .

Run with llama.cpp

# Interactive / one-shot completion (base model = text completion, not chat)
llama-cli -m Qwen3-0.6B-Q4_K_M.gguf -p "The capital of France is" -n 128

# OpenAI-compatible local server
llama-server -m Qwen3-0.6B-Q4_K_M.gguf -c 4096

Run with Python (`llama-cpp-python`)

from llama_cpp import Llama

llm = Llama(model_path="Qwen3-0.6B-Q4_K_M.gguf", n_ctx=4096)
out = llm("The capital of France is", max_tokens=128)
print(out["choices"][0]["text"])

Also loadable directly in Ollama, LM Studio, Jan, and other GGUF runtimes.

Reproduction

The full, scripted pipeline (download → convert → quantize → benchmark → package) is here:

git clone https://github.com/majipa007/LLM-Quantization-Lab
cd LLM-Quantization-Lab
cp config.env.example config.env      # edit if desired
./scripts/00_install_dependencies.sh  # system packages (once)
./run_pipeline.sh                     # runs stages 01–12

Individual stages (build llama.cpp, download, convert, quantize, size/speed/perplexity benchmarks, sample generation, checksums, upload packaging) live under scripts/ and are documented in the repository.

License & attribution

License: Apache-2.0, inherited from the base model.
Base model: Qwen/Qwen3-0.6B-Base by the Qwen team, Alibaba Cloud. All rights to the original weights belong to the authors.
Quantization: produced with llama.cpp (ggml-org/llama.cpp).
These GGUF files are derivative works of the base model and are distributed under the same license. Please cite and credit the original Qwen3 authors.

Integrity

SHA-256 checksums for every file are provided in SHA256SUMS (when included in the upload).

Downloads last month: -

GGUF

Model size

0.8B params

Architecture

qwen3

Hardware compatibility

1-bit

2-bit

3-bit

4-bit

5-bit

6-bit

8-bit

Model tree for Majipa/Qwen3-0.6B-GGUF-Quantized

Base model

Qwen/Qwen3-0.6B-Base

Quantized

(76)

this model

Qwen3-0.6B — GGUF Quantizations

Model details

Conversion & quantization process

Available quantizations & file sizes

Recommended files

Speed benchmarks

Quality evaluation

Hardware specifications (build & reference)

Usage

Download a single file

Run with llama.cpp

Run with Python (llama-cpp-python)

Reproduction

License & attribution

Integrity

Model tree for Majipa/Qwen3-0.6B-GGUF-Quantized

Run with Python (`llama-cpp-python`)