Two days ago we made YuE2-3B the hero: M-A-P’s open song-generation model that plans a piece as editable symbolic notation before rendering audio. The obvious objection was practical. The original weights want serious hardware, which puts them out of reach for most musicians.
That changed quickly, which is the pattern for any open model people actually want to run. audio-cpp/Yue2-3B-GGUF is a quantized conversion for the audio.cpp runtime, and our tracker shows it up 31% in a day. The ComfyUI packaging of the same model rose 142% over the same period. The model is moving from research repo to something people run on their own machines.
What the quantizations cost you
| Build | File size | Peak VRAM | Speed (RTX 5090) |
|---|---|---|---|
| BF16 + F32 VAE | 7.26GB | 12.5GB | 0.27 RTF |
| Q8_0 + F16 VAE | 4.26GB | 8.9GB | 0.20 RTF |
| Q4_0 + F16 VAE | 2.67GB | 7.8GB | 0.20 RTF |
RTF is real-time factor: how long generation takes relative to the length of audio produced. At 0.20 RTF, the model produces roughly five seconds of music per second of compute. The card’s own measurement has Q8_0 rendering 194 seconds of audio in 38.8 seconds. A full-length song in well under a minute.
The important number is the 7.8GB VRAM floor for the 4-bit build. That’s a 3060 or a 4060, not a data-center card, and far below the 24GB the unquantized model wanted. On Apple Silicon, unified memory in that range is ordinary.
Why audio.cpp rather than Python
audio.cpp does for audio models what llama.cpp did for language models: a C++ runtime with quantized weights, a command-line interface, and no Python environment to maintain. We covered the project when it launched. This release is the payoff. The CLI takes a model family, a backend, a style description and a seed, and writes a file.
For creative users, that matters in three specific ways: it runs on ordinary GPUs, it starts fast enough for iteration, and it has no dependency stack to break three weeks after you set it up. A tool you’ll still be able to run next year is worth more than one that’s 5% better today.
The license hasn’t changed
The weights remain CC BY-NC 4.0, inherited from upstream. Non-commercial. Quantizing a model doesn’t relicense it. That’s still the sharpest difference between YuE2 and the commercial services it benchmarks against: you can run it privately, offline, at no per-song cost, and you can’t sell what comes out.
For personal work, sketching, research and installation pieces where nothing is sold, that’s a good deal. For anyone scoring client work, it isn’t, and no amount of local performance changes it.
Note that the GGUF conversion currently tracks the dev branch of audio.cpp for YuE2 support, so expect the usual early-days rough edges.