YuE2
A 3.6B open song generator that writes an editable score first, then renders vocals and accompaniment. CC BY-NC, 24 GB GPU.
In short. YuE2 writes a readable score for a song, then turns it into a recording with singing and instruments, so you can edit the melody or chords before you render. It runs on one 24 GB GPU, and its license is non-commercial.
Summary
M-A-P, an open research community, released YuE2 in September 2026 together with Tokenwave.AI, MBZUAI and ACE Studio. A single model with about 3.6 billion parameters first writes a score in ABC notation with melody and chords, then expands it into semantic music tokens and finally into acoustic features by flow matching. A decoder called YuE2-Vae turns those into 48 kHz stereo audio. Because the plan is a readable score, you can edit the notes, bring your own score, make a cover from a transcription made with M-A-P's SheetSage2, or let an agent revise the score and lyrics between renders.
On the 192 prompts of WildSongBench, the YuE2 team's own benchmark, YuE2 scores 6.73 on the SongBench average (6.96 when the best of eight is chosen). In the same table Suno v5 gets 6.87, MiniMax Music 3 gets 6.28, HeartMuLa 6.25 and ACE-Step 1.5 6.01. The authors say YuE2 is competitive with Suno v5 and v6. The numbers come from the YuE2 team, and the YuE2 scores use its own candidate selection (the better of two songs, or the best of eight). A paper on arXiv (2609.33757) describes the method.
The card asks for Linux, Python 3.10 or later, an NVIDIA GPU with BF16 support and 24 GB of memory, and it measures 11 to 14 GiB of peak VRAM. A 3.6-minute song took 71 seconds on an RTX 4090 in its tests. The weights are CC BY-NC 4.0, so any commercial use needs a separate license from the authors. In the 30 days to 11 November 2026, Hugging Face counted about 22,500 downloads of the official repository, and a community GGUF conversion had about 250,000.
Good for
- Songs whose melody and chords you want to edit before the audio is rendered
- Cover versions of existing songs, with the SheetSage2 transcriber
- Research on score-conditioned music generation, with an arXiv paper and an open benchmark
- Local generation of full songs on one 24 GB GPU
Watch out for
- The weights are non-commercial only, so selling the songs or building a paid service needs permission from the authors
- Linux and an NVIDIA GPU with BF16 support are required, and the card recommends 24 GB of GPU memory and 24 GB of RAM
- The benchmark is the YuE2 team's own, and cover and agent editing need extra tools and some setup
- The card does not describe the training data, so ask a lawyer about copyright before you use the output commercially
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of YuE2. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| YuE2-3B (BF16, with YuE2-Vae) BF16 | 3.63B | 7.8 GB | 24 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 24 GB of GPU memory.
The license in plain words
A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0.
- Attribution or model name must be shown
See also
- Best open-source music and sound-generation models in 2026
- Open music models grew up in 2026: seven models for songs, instruments and sound effects
Frequently asked questions
- What is YuE2?
- YuE2 writes a readable score for a song, then turns it into a recording with singing and instruments, so you can edit the melody or chords before you render. It runs on one 24 GB GPU, and its license is non-commercial.
- Can I use YuE2 commercially?
- Not without separate permission. YuE2 is released under the CC BY-NC, which allows research and personal use only.
- Is YuE2 open source?
- YuE2 is open-weight, not open source in the strict sense. You can download the model files, but the CC BY-NC adds conditions that the Open Source Initiative definition does not allow.
- Can I run YuE2 on my own computer?
- Yes, on a good graphics card. The smallest variant, YuE2-3B (BF16, with YuE2-Vae), needs about 24 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was YuE2 released?
- YuE2 was released by M-A-P on September 9, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- YuE2-3B model card (Hugging Face) huggingface.co
- YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality (arXiv) arxiv.org
- YuE2 project page: demos, agentic editing and benchmark (M-A-P) map-yue2.github.io
- YuE2-3B license, CC BY-NC 4.0 (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| ACE-Step 1.5 ACE-Step | 2B – 4B | – | MIT (Permissive) | 4 GB Laptop or phone | |
| MiniMax Music 3 MiniMax | 11.1B | – | MiniMax Music3 Community (Conditional) | 24 GB Gaming GPU | |
| HeartMuLa HeartMuLa | 3.94B | – | Apache 2.0 (Permissive) | 26 GB Gaming GPU | |
| Stable Audio 3 Stability AI | 460M – 1.4B | – | Stability Community (Conditional) | 4.02 GB Laptop or phone |