Score-first composition
The model writes a melody-and-chord plan (ABC notation) before it sings. Keep it, edit it, or paste your own score.
Weavy Studio turns your lyrics into a finished, mastered-ready song with the open YuE2 model — and hands you every dial, from the score plan to the sampler.
Open the studioThe model writes a melody-and-chord plan (ABC notation) before it sings. Keep it, edit it, or paste your own score.
Style strength (CFG), seed, length and temperature, top-p, top-k and repetition penalty for both the score and the audio stage.
Render up to 8 versions in parallel on separate GPUs and keep the best one — best-of-N beats a single roll.
48 kHz stereo FLAC (24-bit) for mastering plus a 320 kbps MP3 for listening, straight from the render.
Watch each take move from planning to generating to encoding in real time — no refreshing, no polling.
Everything the studio does is one HTTP call away: create, list, cancel and download generations with an API key.
YuE2-3B by the Multimodal Art Projection team (M-A-P), an open-weight lyrics-to-song model released in September 2026, decoded to 48 kHz stereo by its YuE2 VAE.
English sings best. Many other languages work — Polish, Spanish, German, French, Portuguese and more; put the language first in the style.
About 1.5 minutes of GPU time for a 3–4 minute song. Several takes render in parallel.
The YuE2 weights are licensed CC BY-NC 4.0 with extra permission for individual creators to monetise their outputs; companies need a commercial licence from M-A-P. Check the licence for your case.
Yes — a documented REST API with API keys. See the API reference.