YuE2-3B Beats Suno v5 on Its Own Music Benchmark

Multimodal Art Projection, a research group known as m-a-p, released an open-weight AI model Sept. 9 that turns lyrics into a finished song. The model, YuE2-3B, beat Suno's v5 on m-a-p's own benchmark, though its weights ship under a license that bars commercial use.

YuE2-3B has 3 billion parameters and builds a song in two steps. First, what m-a-p calls an AR-NAR Mixture-of-Transformers backbone, one model predicting tokens in sequence and a second filling in the rest in parallel, writes an editable musical score in ABC notation, laying out melody and chords. Then a flow-matching model renders that plan into 48 kHz stereo audio, according to the model's card on Hugging Face.

How YuE2-3B compares with Suno, by m-a-p's own account

M-a-p ran the comparison itself, using 192 prompts from a test it calls WildSongBench, and no independent lab has repeated the evaluation. The win only shows up when the model samples eight times and keeps the best result; run once, YuE2-3B scored 6.73, behind the 6.94 that AlphaSignal recorded for Mureka 9, a proprietary model it listed as the current single-generation leader.

Model	Mode	SongBench average
YuE2-3B	Best of 8	6.9632
YuE2-3B	Single generation	6.73
Suno v5	Single generation	6.8721
Suno v6	Single generation	6.5562
Mureka 9	Single generation	6.94

What it takes to run: one consumer GPU

YuE2-3B runs on a single 24-gigabyte GPU without quantization, m-a-p said, generating a 3.6-minute song in about 71 seconds on an Nvidia RTX 4090 and using a peak of 11.18 gigabytes of video memory. On Nvidia's H800 chip, the same job used 10.34 gigabytes. That is a far lighter hardware bar than DeepSeek's 552-billion-parameter V4.1-Flash release, which needs a multi-GPU server rather than a single desktop card.

The license: free to use, not free to sell

M-a-p released YuE2-3B's weights under a Creative Commons BY-NC 4.0 license, which permits personal and research use but not commercial products, while the inference code carries a separate Apache 2.0 license. That split means a company cannot ship YuE2-3B inside a paid product without separate permission, even though the code that runs it is unrestricted.

Built on a 2025 predecessor

YuE2-3B extends a project m-a-p first published in a March 2025 paper on arXiv titled "YuE: Scaling Open Foundation Models for Long-Form Music Generation." The new version adds the editable-score step and what the card calls agentic editing, where another AI model revises the score, style or lyrics after a first pass, rather than restarting generation from scratch. That editability is the feature m-a-p is betting on: Suno and Mureka both return finished audio with no intermediate layer a musician can open and change.

It joins a run of small, open-weight releases this month that includes OpenBMB's 2.5-billion-parameter MiniCPM5. Native support for YuE2-3B landed in ComfyUI's master branch Sept. 11, two days after the model's release, according to ComfyUI's own changelog, letting users chain it into existing image and video generation workflows without custom code.

M-a-p has not said whether a larger YuE2 checkpoint is coming, or whether a future version will carry a license that allows commercial use.