echomom/echomom-minimax-music3-q8

Echomom MiniMax Music3 Q8 + BF16 Depth Decoder

A curated mixed-precision inference pack for ServeurpersoCom/minimaxmusic.cpp, prepared for the Echomom RunPod experiment and tested around commit 381c9b7. The language model and Flow DiT transformer use Q8_0, while the RVQ depth decoder uses BF16.

The converted weights come from Serveurperso/MiniMax-Music3-GGUF at revision 9cdffedb54de2509ae55a6831a677645fb353a7d. The original model is MiniMaxAI/MiniMax-Music3.

Components

Component	Format	File
Language model	Q8_0	MiniMax-Music3-language_model-Q8_0.gguf
RVQ depth decoder	BF16	MiniMax-Music3-rvq_depth_decoder-BF16.gguf
Flow DiT transformer	Q8_0	MiniMax-Music3-transformer-Q8_0.gguf
Condition encoder	F32	MiniMax-Music3-condition_encoder-F32.gguf
Vocoder	F32	MiniMax-Music3-vocoder-F32.gguf

The five weight files total 13,428,487,488 bytes (approximately 13.43 GB or 12.51 GiB). Actual resident VRAM depends on runtime configuration and allocator overhead.

License

The included LICENSE is copied from the original model at revision fbdf52fbaaca799592917417eb05f1899f1255ec. The MiniMax Community License in that file governs the weights. Review it before using or distributing this model.