| --- |
| library_name: audio.cpp |
| pipeline_tag: text-to-audio |
| license: other |
| license_name: minimax-music3-community-license |
| base_model: |
| - MiniMaxAI/MiniMax-Music3 |
| tags: |
| - audio.cpp |
| - gguf |
| - minimax-music3 |
| --- |
| |
| # MiniMax Music 3 GGUF |
|
|
| GGUF package for MiniMax Music 3 for audio.cpp. |
|
|
| Upstream model: https://huggingface.co/MiniMaxAI/MiniMax-Music3 |
| Upstream license: https://huggingface.co/MiniMaxAI/MiniMax-Music3/blob/main/LICENSE |
|
|
| ## Preview Notes |
|
|
| - The implementation is available on the `preview/minimax-music-3` branch while the runtime surface is iterated quickly. |
| - The current runtime uses model-local resource loading instead of treating the v1 spec as the runtime contract. This keeps component selection flexible while the package layout and option surface settle. |
| - The default component mix now favors Q4_0 for the large language model and flow transformer, with BF16 kept for the RVQ depth decoder. |
| - BF16, Q4_K, and Q4_0 component variants are included for quality/performance comparison. The best package-wide component mix is still being measured. |
| - Longer generations such as five-minute songs are supported as long-form runs, but they are currently tuned for completion and quality checks rather than realtime throughput. |
| - Memory usage remains an active optimization target for larger durations and alternate component mixes. |
| |
| ## Quick Start |
| |
| ```bash |
| audiocpp_cli \ |
| --task gen \ |
| --family minimax_music3 \ |
| --model MiniMax-Music3-GGUF \ |
| --backend cuda \ |
| --text "A bright pop rock song with clean drums and a clear male vocal." \ |
| --request-option 'lyrics=[verse] City lights are shining low. I keep moving with the glow. [chorus] Turn it up and let it fly. Sing the melody tonight.' \ |
| --request-option duration_sec=20 \ |
| --request-option num_inference_steps=30 \ |
| --out output.wav |
| ``` |
| |
| ## Components |
| |
| Default audio.cpp component mix: |
| |
| - `language_model_q4_0.gguf` |
| - `rvq_depth_decoder_bf16.gguf` |
| - `transformer_q4_0.gguf` |
| - `condition_encoder.gguf` |
| - `vocoder.gguf` |
| |
| The BF16, Q4_0, and Q4_K component variants are included for measurement and quality/performance comparison. |
| |
| Component GGUFs can be selected explicitly for experiments: |
| |
| ```bash |
| --session-option minimax_music3.language_model_gguf=language_model_bf16.gguf |
| --session-option minimax_music3.language_model_gguf=language_model_q4_k.gguf |
| --session-option minimax_music3.rvq_depth_decoder_gguf=rvq_depth_decoder_q4_k.gguf |
| --session-option minimax_music3.flow_transformer_gguf=transformer_bf16.gguf |
| --session-option minimax_music3.flow_transformer_gguf=transformer_q4_k.gguf |
| ``` |
| |
| ## License |
| |
| This GGUF package follows the upstream MiniMax-Music3 COMMUNITY LICENSE. The |
| license text is included in `LICENSE`; review it before use, especially for |
| commercial deployment. |
| |
