open-qwen-music-weights

Open-Qwen-Music

An independent research reproduction of the Qwen-Music method for text-to-music generation

GitHub · Dataset · Demo · Reference Paper

Open-Qwen-Music is an independent research reproduction based on the architecture and training methods described in the Qwen-Music Technical Report. It is not an official Qwen release and is not affiliated with, endorsed by, or maintained by the Qwen team or Alibaba Cloud.

The checkpoints in this repository were prepared by this project. They are not official Qwen-Music checkpoints or evaluation artifacts.

Open-Qwen-Music generates 48 kHz stereo music from text descriptions, musical tags, lyrics, and section-level structure. A music language model first produces 25 Hz semantic tokens, which are converted to audio by an acoustic renderer, spectrogram VAE, and bandwidth refiner.

Model Provenance

The bundle contains no official Qwen-Music checkpoint. The renderer also uses the pinned Qwen/Qwen3-Embedding-0.6B text encoder at runtime; that external model is downloaded separately and is not included in this repository.

Model Components

The semantic interface uses one 32,768-entry codebook at 25 Hz. The released language model uses a plain semantic-token sequence and was trained for 10,000 optimizer updates in two 5,000-update phases.

Installation

Clone the project source and install the inference dependencies:

git clone https://github.com/biang15343100-source/Open-Qwen-Music.git
cd Open-Qwen-Music
python -m venv .venv
source .venv/bin/activate
pip install -e ".[render]"

Generate Music

Run the complete text-to-music pipeline with the included prompt:

python scripts/infer.py \
  --model oqmtest1451/open-qwen-music-weights \
  --prompts examples/prompts.jsonl \
  --output-dir outputs/demo \
  --device cuda

The default language-model sampling temperature is 0.7. Each prompt produces a 48 kHz stereo WAV file and a JSON metadata sidecar. See the GitHub inference guide for the prompt schema, offline download instructions, and runtime options.

To download the bundle in advance:

hf download oqmtest1451/open-qwen-music-weights \
  --local-dir checkpoints/open-qwen-music
python scripts/infer.py \
  --model checkpoints/open-qwen-music \
  --prompts examples/prompts.jsonl \
  --output-dir outputs/demo \
  --device cuda

Reference Method

This implementation is inspired by the architecture and training methodology described in the Qwen-Music Technical Report by Xu et al. The paper is cited as the source of the method. Its reported experiments and results belong to the original authors and should not be interpreted as results of this reproduction.

This project uses independently prepared data, training infrastructure, checkpoints, and demonstrations. Differences in data, scale, optimization, and implementation may produce results that differ from the original report.

Training Data

The training corpus is published as david-miller-45678/open-qwen-music-dataset-c-minimax-music3. The model bundle is tied to dataset revision 454294f27d920bb3c1430566bb7f5c153ece4c27, containing 62,417 audio records and structured annotations in WebDataset format.

The dataset is restricted to non-commercial academic research and includes source-specific terms. Review its LICENSE, NOTICE, source policy, and per-record rights metadata before use.

Hardware

End-to-end inference loads the language model, acoustic renderer, VAE, refiner, and text encoder. A CUDA GPU with bfloat16 support is recommended. CPU execution is intended for interface checks and is not practical for text-to-music generation.

Model Layout

language-model/       Music language model and text tokenizer
semantic-tokenizer/   Audio-to-semantic tokenizer
acoustic-renderer/    Acoustic diffusion transformer
acoustic-vae/         Spectrogram autoencoder
bandwidth-refiner/    High-frequency refinement model
assets/                Latent statistics and semantic embeddings

Limitations

This is a research reproduction, and it should not be expected to match the original Qwen-Music system. Generated music can contain structural drift, imperfect lyric intelligibility, unstable long-range repetition, and audio artifacts. Outputs may reflect biases in the training data. Review generated audio before publication or downstream use.

License and Attribution

This repository contains components with different provenance.

The language-model component is derived from Qwen/Qwen2.5-Omni-3B and remains subject to the included Qwen Research License. The required Qwen license and attribution notice are retained in this repository.

The other checkpoints were trained by the Open-Qwen-Music project. Their terms are described in the repository LICENSE. Third-party code and architecture adaptations retain their respective notices.

The use of the name “Qwen” identifies the referenced method and upstream model lineage. It does not indicate affiliation with or endorsement by the Qwen team.