File size: 2,845 Bytes
da36da5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ead13fe
 
c36aaee
 
39bf675
c36aaee
39bf675
 
c36aaee
39bf675
c36aaee
da36da5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39bf675
da36da5
39bf675
da36da5
 
 
39bf675
da36da5
 
 
 
 
39bf675
da36da5
 
39bf675
da36da5
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
---
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

Compatible Q8 GGUF package: https://huggingface.co/joemattie/MiniMax-Music3-GGUF

## 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.