03 — Inference · systems/Shipped
MiniMax Music 3, local
The official guidance was two GPUs. I got it onto a single consumer card, and faster than the path it shipped with.
- Speed-up
- 3.2×
- 20s song in
- 43.1s
- Peak VRAM
- ~9GiB
- Card
- RTX 3060 12GB
Role
Inference engineering, CUDA kernels
Year
2026
Stack
- C++ / CUDA
- audio.cpp
- RTX 3060
- 11B weights
The full 11B stack on one RTX 3060
What happened
MiniMax Music 3 came out as an eleven billion parameter stack with a recommended setup most people at home do not have. I wanted the whole thing on one card — not a distilled version, not a quantised approximation that loses the top end.
It runs in about nine gigabytes and turns out a twenty second song in forty-three seconds, which is 3.2× quicker than the official path on the same hardware.
Where the time went
The gains are a stack of small wins, each measured. Slicing the vocabulary down to reachable rows took 5.6 seconds off the lm_head. A second-order Adams-Bashforth flow solver took another 8.5. Bucketing cache growth saved 9.5 on longer songs. Fusing QKV in the depth decoder, routing MMVQ through the MMQ tensor-core path, and dropping 213,846 bf16 parity casts took the rest.
A few of the changes were correctness, not speed: graph inputs moved into dedicated buffers to stop silent corruption, and an F16 KV cache that halved the allocation and bought about 50% more song length.
Where it lands
A 75-second song renders in 171 seconds, which is 2.28× realtime. A 2:24 track takes 361. Flow matching runs at roughly 79% of the bf16 FLOP ceiling and the autoregressive stages at 71% and 55% of memory bandwidth — there is not a lot of headroom left on this card.
Why the number matters
A model that needs a cluster is a demo. A model that runs on the card already in your machine is infrastructure. That gap is where the entire argument about open music AI sits, and it is closed by memory work and scheduling, not by a bigger budget.
Everything I learn here ends up in the studios — ACE-Step UI and HeartMuLa Studio both exist to put this class of model in front of somebody who just wants to write a song.
