YuE2 Studio generates full songs on your own GPU, and lets you edit the score before it sings

Paid AI song services keep the model, the score and the master. YuE2 Studio, released this week on GitHub, hands all three back and runs on the card already in your PC.

The project wraps YuE2, the open song model from M-A-P, in a native Windows app built with Tauri, Rust and React. The repository was created on 2026-09-22 by developer timoncool and reached 241 stars and 23 forks in six days, under an MIT licence. It ships as one Windows x64 executable with an auto-updater and asks for an NVIDIA card with 6 GB of VRAM or more. AMD and Intel cards work through Vulkan, and timoncool marks that path experimental after hearing garbled vocals on AMD Radeon integrated graphics.

What the model writes before it sings

YuE2 writes a composition in ABC notation first, then performs it. YuE2 Studio engraves that ABC as sheet music, so a user can read the melody and chords, change the notes, tempo or key, and render the same composition again with a different voice or instrument. Songs run up to six minutes in the languages the model sings. The README states that a 3:38 song renders in about 46 seconds on an RTX 4090 with the Q8_0 model set. Both figures are the developer's own on that specific hardware.

There is also a covers mode. A companion component, SheetSage2, listens to a recording and writes its melody down; YuE2 then sings that melody in the style you pick. timoncool notes the lyric has to fit the melody line by line, or the singing drifts, and transcription mistakes carry into the cover.

Every result is a file on your disk

The library sits in Windows folders you choose, and every save uses the platform's own Save dialog. Songs export as MP3 or as 16, 24 or 32-bit WAV. Stems, MIDI conversions, lyric sheets and scores each become tracks in the same library, tagged with the settings they were made from. Prompts save as JSON or YAML in the engine's request format, so they carry over to the yue2.cpp WebUI or the command-line yue-synth unchanged. Every track keeps its request and its audio codes for bit-for-bit replay.

The studio also exposes an MCP server. An agent can connect to http://127.0.0.1:8791/mcp and drive the studio from Claude Code or any MCP-compatible client, reading a skill file the studio serves on the same port.

Suno and Udio are the paid services this kind of tool competes with; both charge a monthly subscription and both keep every render on their servers. YuE2 Studio puts the master on your disk and the composition in your hands. Being able to edit the melody as sheet music before the vocal is generated, and to keep the same tune with a different performance, is a musician's control that Suno and Udio do not expose today.

For a marketing team producing jingles or explainer soundtracks, the trade is a Windows machine with a modern GPU and slower renders than a cloud service, in exchange for no subscription and audio that stays in the office. The MCP endpoint means song rendering fits into the same agent pipeline as text and image generation.

Source

YuE2 Studio on GitHub, by timoncool. Model: YuE2 from M-A-P. Native engine: yue2.cpp.