infosave/minimax-music3-cortiq

MiniMax-Music-3, running on one Rust binary

Caption and lyrics in, 44.1 kHz stereo out — from a single 5.55 GB .cmf read by `cortiq`. No PyTorch, no diffusers, no Python anywhere in the inference path: the app here only fetches the weights, shells out to the binary and shows a waveform.

Weights: infosave/MiniMax-Music-3-cmf — MiniMax's 20.3 GB of originals packed into 5.55 GB, cortiq verify clean, tokenizer inside the file.

What runs when you press the button

Three stacks in sequence, all out of that one file:

1. 1.The AR stack generates the conditioning rather than encoding it — a Qwen3-8B backbone prefilled at batch two for classifier-free guidance, then sampled one audio frame at a time at 25 fps, with seven more codebooks per frame through an RVQ depth decoder.
2. 2.A flow-matching DiT denoises the latent, Euler, windowed 689 frames at a time.
3. 3.The DAV vocoder turns each latent frame into 512 stereo samples.

On the hardware

This Space is CPU-bound and slow — a four-second clip is minutes, not seconds, and the first request also pays the 5.55 GB download. The sliders are capped where they are for that reason, and the estimate above the button recalibrates itself from your first real run rather than guessing.

The same command locally, where a GPU is available:

cargo install cortiq-cli
hf download infosave/MiniMax-Music-3-cmf minimax-music3-q4tp.cmf --local-dir .
cortiq music minimax-music3-q4tp.cmf \
  --prompt "bpm is 92, key is E minor. Electric blues rock, gritty slide guitar." \
  --lyrics "[verse]
I woke up on a dusty road" \
  --seconds 15 --steps 8 --seed 42 --out song.wav

Same seed, same prompt, same song.