[audio.cpp] C++/GGML MiniMax-Music3. Fits comfortably within 16 GB VRAM and supports CUDA, Vulkan, Metal, and AMD GPUs. Demo included.

[0xShug0]

Repo: https://github.com/0xShug0/audio.cpp

Measured on an RTX 5090 with CUDA using a 30-second lyric generation request, 30 flow steps, flow guidance scale 1.7, AR guidance scale 1.5, and top-k 50. Peak VRAM is the observed nvidia-smi process peak during a warmup-plus-measured-request run.

Component mix	Language model	RVQ depth decoder	Flow transformer	RTF	Speed	Peak VRAM
Default Q4/Q8/Q4	q4_0	q8_0	q4_0	0.738	1.35x faster realtime	9.8 GiB
Q8	q8_0	q8_0	q8_0	0.832	1.20x faster realtime	13.4 GiB
BF16	bf16	bf16	bf16	1.389	0.72x realtime	19.4 GiB

60s and 180s (target duration) end to end demo, using the features/pop prompt from official demo

music3_demo.mp4