Open music generation has long had an awkward gap: text and image models already live in a generate-revise-regenerate loop, while song models still work like pulling a slot-machine lever. On September 10, the music generation model YuE2-3B landed on Hugging Face (model card); according to the open-weights tracker The Open Weights, it comes from OpenMOSS, and its whole design attacks that pain point: write the score first, then sing it.
The score becomes a first-class citizen
The backbone is an AR–NAR Mixture-of-Transformers: the model first emits a score written in ABC notation plus semantic tokens, then produces acoustic latents through flow matching, and a VAE decodes them into 48 kHz stereo audio. Planning and synthesis are also separated at the API level — pipe.plan() returns the score, you can edit melody and chords by hand, then hand it back to the pipeline for rendering. Three control levels are offered: melody plus chords, melody-only, and fully free generation. Covers follow the same road: transcribe the original song into an ABC score with SheetSage2, align lyrics with tools like Qwen3-ASR, then re-render in a new style.
The model card also demos an agentic-editing case of 9 steps and 14 versions: feedback like "make it jazz, add a sax solo" goes to an agent that revises the score, style and lyrics while YuE2 renders the next version — music generation turns from a one-shot gamble into a conversational collaboration.
A song in 71 seconds on a consumer GPU
With a 3B-class parameter count, the deployment bar drops accordingly: official tests show a 3.6-minute song generated in 71 seconds on an RTX 4090, peak VRAM around 11.2 GiB, no quantization needed; a single 24GB GPU plus 24GB of host RAM covers the full create-cover-edit workflow. On the server side, with vLLM 0.19 at 32-way concurrency the system sustains 3,231 tokens/s and roughly 373 songs per hour.
The report card on an official benchmark
On the team's own WildSongBench (192 prompts), YuE2 best-of-8 reaches a SongBench average of 6.9632, above every open and proprietary system in the table — Suno v5 sits at 6.8721, and open-source rivals LeVo 2 (6.3247) and MiniMax Music 3 (6.2830) trail well behind; on the zero-shot cover task SHS100K, CLEWS Hit@1 hits 71.3%. One caveat: these are official numbers on a self-built benchmark, the technical report is marked "coming soon," and the figures deserve independent replication before being treated as settled (tracker coverage).
The weights ship under CC BY-NC 4.0 — fine for research and personal use, off limits for commercial deployment; the model card shows 19 downloads in the last 30 days, so it has only just entered the community's field of view.
For people who make music, the real takeaway is not the benchmark score but the fact that the score itself — the source code of the music industry — is handed back to the user: editing two bars and re-rendering beats re-rolling an entire song. And a 3B-class model squeezing full song generation onto a consumer GPU is one more sign that vertical modalities may not need frontier-model scale. What to watch next: the formal technical report and what the community does with it.