Qwen3-0.6B-GGUF-Quantized

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

Conversion & quantization process

1. Download the source model from the Hugging Face Hub.
2. Convert it to a full-precision BF16 GGUF with llama.cpp's convert_hf_to_gguf.py.
3. 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.

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).

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.

Hardware specifications (build & reference)

Quantization and (upcoming) benchmarks were produced on:

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).