MiniMax Music 3

MiniMax Music 3 pairs an autoregressive semantic language model with an RVQ depth decoder and a flow matching transformer. It gives expressive vocal performances and dynamic song structures.

It is also the largest model Miso installs, at around 13 GB.

What it does in Miso

One task, generate.minimax, on the create page, using the guided prompt builder.

Vocals

Singing with strong dynamic range. Lyrics are required.

Options

Option	Default	Does
text		A detailed caption: genre, production, instruments, vocal timbre
lyrics		Formatted with bracket tags
duration_sec	20	Frame budget in seconds. Larger values allow longer output and use more memory
num_inference_steps	30	Flow matching steps
guidance_scale	1.7	Flow transformer CFG scale
ar_guidance_scale	1.5	Autoregressive CFG scale
top_k	50	Token sampling limit

Prompting

MiniMax does best when the prompt reads like a production brief rather than a list of genres.

A melodic indie rock song with clean chime guitars, warm bass,
punchy room drums, clear female lead vocals, upbeat tempo,
clean studio recording.

Lyrics want lowercase bracket tags:

[verse]
Morning sun breaks through the bedroom blinds
Coffee in a chipped ceramic cup

[chorus]
Run until the road turns into sand
Feel the wind against our open hands

[outro]
All to find... fading out.

Memory

MiniMax uses component GGUFs: a language model, an RVQ depth decoder, a condition encoder, a flow transformer, and a vocoder.

Miso sends minimax_music3.mem_saver=true, which loads the stages one after another and fits the run into 12 to 16 GB of VRAM.

Raising duration_sec raises memory use. If a long generation fails on a 16 GB card, bring it down first before changing anything else.