What is YuE2-3B?
YuE2-3B is an open music model from the Multimodal Art Projection (M-A-P) research community, the group behind the original YuE. It turns a style prompt and lyrics into a full song with vocals and accompaniment in 48 kHz stereo. Before rendering audio it writes a symbolic plan, an ABC score with melody and chords, which you can export, edit by hand or with an AI agent, and feed back in.
The weights went up on Hugging Face in September 2026 under CC BY-NC 4.0, which forbids commercial use; the code repository on GitHub is Apache-2.0. The model card's demos run up to 5:00, in English and Mandarin. It also makes covers: transcribe an existing song's melody with the companion SheetSage2 model, and YuE2 sings it in a new style.
Is YuE2-3B worth it?
YuE2-3B makes the best-scoring songs of any open model on paper, and you cannot sell them. On WildSongBench, its authors' 192-prompt benchmark, YuE2 picking the best of eight takes scored a SongBench average of 6.9632, ahead of every proprietary model tested, including Suno v5 at 6.8721 and Mureka 9 at 6.9377. Choosing between two takes, it scored 6.7316, still above MiniMax Music 3 (6.2830) and ACE-Step 1.5 (6.0118). These are the authors' figures on their own benchmark.
Its distinctive feature is the score. YuE2 first writes melody and chords in ABC notation, then generates audio from that plan. You can stop after planning, change a chord or a phrase, and render again with the same seed, so a revision changes what you asked for rather than the whole song. The demo shows an agent taking one song through 9 steps and 14 versions. The same mechanism makes covers that keep the original tune: given the source score, it identified the original song 71.3% of the time on the SHS100K test, against 2.4% for ACE-Step 1.5.
It is practical on one consumer card. On an RTX 4090 a 3.6-minute song takes 71 seconds with 11.18 GiB of peak VRAM; the card asks for a 24 GB GPU and 24 GB of free system RAM. A separate vLLM server on one H800 reached 373.53 songs an hour with 32 concurrent requests.
The catch: CC BY-NC 4.0 rules out ads, games, apps and anything else that earns money, whoever runs the weights. The quick start is Linux and NVIDIA only, with BF16 support, and the card lists only English and Mandarin. Lyrics are clear but not the clearest: its phoneme error rate of 8.44% is above MiniMax Music 3's 6.27%.
Our take: for researchers, hobbyists and anyone writing songs for themselves, YuE2's editable score makes it the most controllable open option; once a track is headed for a client or a store, the licence sends you to MiniMax Music 3 or ACE-Step 1.5.
Where it earns its place
A university music-technology lab
Compare AI song generators and study controllable composition
Export YuE2's ABC plans, alter harmony or melody under controlled conditions and render again with fixed seeds; the CC BY-NC licence covers academic research, and the WildSongBench prompts are public on Hugging Face.
A hobbyist songwriter with an RTX 4090
Hear demo versions of their own lyrics in different genres
Run the pipeline with cot="full", edit score.abc to fix a chorus melody and render again; a song of three to four minutes takes about a minute, and nothing leaves the machine.
A choir director preparing rehearsal material
Hear a public-domain carol in a new arrangement
Transcribe the melody with SheetSage2, supply it with cot="melody" and a new style prompt; fine for rehearsal and teaching, not for a paid recording.
Before you commit
- Setup
- On Linux with Python 3.10+, install the yue2_infer wheel from the Hugging Face repo and load YuE2Pipeline.from_pretrained("m-a-p/YuE2-3B"); the default YuE2-Vae decoder is a separate download.
- Integrations
- Python pipeline with separate plan, semantic, synthesis and decode stages; a yue2 command-line tool with batch mode; a vLLM-based server for concurrent requests; SheetSage2 for melody transcription.
- Data and privacy
- Runs locally; prompts and audio stay on your machine. The model card does not describe the training data.
- Cost at scale
- No licence fee, but commercial use is not allowed at any scale. The card reports 373.53 songs an hour on one H800 at 32 concurrent requests.
- Limits
- One song at a time in the Hugging Face pipeline; needs a 24 GB GPU and 24 GB host RAM; Linux and NVIDIA with BF16; covers need a separately transcribed melody and lyrics.
What’s good
- Highest SongBench average on WildSongBench (6.9632 best-of-8), per its authors
- Editable ABC score: change melody or chords and render again
- Covers that keep the original melody (71.3% identification on SHS100K)
- 48 kHz stereo; a 3.6-minute song in 71 seconds on an RTX 4090
- Peak VRAM about 11 GiB in the card's tests
What’s not
- CC BY-NC 4.0: no commercial use
- Linux and NVIDIA GPUs with BF16 only
- Benchmarks run by its own authors
- Only English and Mandarin listed
- No hosted API
Best for
Researchers, hobbyists and songwriters working non-commercially who want to revise a song's melody and chords rather than regenerate it.
YuE2-3B alternatives
YuE2-3B: common questions
- Is YuE2-3B worth it?
- YuE2-3B is the open song model to study and tinker with, because you can edit its score note by note, but its non-commercial licence keeps it out of anything you sell. We rate it 4.0 out of 5.
- How much does YuE2-3B cost?
- Free weights under CC BY-NC 4.0: non-commercial use only. No hosted API; you pay for your own GPU.
- What are the downsides of YuE2-3B?
- CC BY-NC 4.0: no commercial use; linux and NVIDIA GPUs with BF16 only; benchmarks run by its own authors; only English and Mandarin listed; no hosted API.
- Who is YuE2-3B best for?
- Researchers, hobbyists and songwriters working non-commercially who want to revise a song's melody and chords rather than regenerate it.
- How much does YuE2-3B cost for a team or in production?
- No licence fee, but commercial use is not allowed at any scale. The card reports 373.53 songs an hour on one H800 at 32 concurrent requests.
- Is YuE2-3B safe for company data?
- Runs locally; prompts and audio stay on your machine. The model card does not describe the training data.
- What are the best YuE2-3B alternatives?
- The closest alternatives we have reviewed are MiniMax Music 3, ACE-Step 1.5 and Lyria 3.5.
Answers checked Nov 7, 2026.
Make YuE2-3B? Show this rating on your site
