# YuE2 Studio YuE2 Studio is an open-source Windows desktop app for YuE2, the song model from M-A-P (the team behind [[YuE]]). You give it a style and lyrics. The model first writes the song as a score (melody and chords in ABC notation), then sings it as a full track with vocals. You get the score back with the audio, so you can change a note, the tempo or the key and render again; the composition stays, only the performance changes. When a music model returns only audio, fixing one melody line means generating the whole song again. Here the composition is plain text you can edit by hand, and an LLM can read and write it too. timoncool (Nerual Dreming) maintains it on GitHub, under the MIT license. The repo was created on 2026-09-22. By 2026-10-04 it had about 440 stars, 39 forks, 22 releases (v1.0.0 to v3.4.0 in nine days) and roughly 25,700 release downloads. ## How it's built - Engine: yue2.cpp by Serveurperso, a native C++/CUDA port of YuE2 on [[Georgi Gerganov Machine Learning (GGML)|ggml]] (the tensor library under [[llama.cpp]]), running quantized [[GPT-Generated Unified Format (GGUF)|GGUF]] weights - App: [[Rust]] (an Axum service on `127.0.0.1:8791`) and [[React]] in a [[Tauri]] window. One executable; no Python, no Node.js, no launcher - CUDA on NVIDIA is the fast, tested path. Vulkan on AMD and Intel is experimental: on AMD Radeon integrated graphics the vocals came out unintelligible. Without a GPU it falls back to the CPU, which works but is many times slower - Models, songs, settings and logs all live in the app's folder, so deleting the folder removes everything. The installed version updates itself YuE2 is a 3B backbone (it writes the score and the audio codes) plus a VAE that turns them into 48 kHz stereo. The original YuE needed a 7B stage, a 1B stage and a 24 GB GPU for 30-second segments. A smaller model and [[AI Quantization|quantization]] bring it down to 6 GB cards: | GPU VRAM | Model set | Download | |---|---|---| | 12 GB and above | BF16, original weights | 9.8 GB | | 8 GB and above | Q8_0, near lossless | 5.1 GB | | 7 GB and above | Q6_K | 4.1 GB | | 5.5 GB and above | Q5_K_M | 3.8 GB | Speed, per the README: a 3:38 song in about 46 seconds on an RTX 4090 with the Q8_0 set. The original YuE took ~360 seconds per 30 seconds of audio on the same card. Different model and different measurement, so it's not a clean benchmark, but per second of audio it's over 50 times faster. ## What it does - Full songs up to six minutes from a style and lyrics - Score only: generate the score, fix it, then render it - Covers: SheetSage2 transcribes a recording's melody into a score, and YuE2 sings that melody in your style. New lyrics need the same number of syllables per line, or the singing drifts off the tune - Every track keeps its request and audio codes, so you can re-render it bit for bit, or with other steps, seeds or output formats, without composing again - Post-production: six stems with HT-Demucs, any track to MIDI with MuScriptor, a MIDI editor (signal), word-level karaoke timed by [[Whisper]] or Parakeet, noise reduction, [[Virtual Studio Technology (VST)|VST3]] plugins, mastering to a reference track, music videos - [[Low Rank Adapter (LoRA)|LoRA]] support: load them, search [[HuggingFace|Hugging Face]] for them, or train your own from a folder of songs (NVIDIA, 11 GB VRAM or more) - A local [[Gemma]] model, [[OpenRouter]] or your connected agent can write the style and lyrics from a one-line idea - A Winamp 2 mode with skinned windows, the classic equalizer and a MilkDrop visualiser (via Webamp and Butterchurn). Useless, and I love it (no kidding!) ## Built for AI agents While it's open, the app serves an MCP server at `http://127.0.0.1:8791/mcp` with 162 tools. An agent such as [[Claude Code]] can do everything the UI does (songs, scores, library, covers, stems, MIDI, LoRA training), and it can also screenshot the window, click and type in it. It also tells the agent how to use those tools: - The server exposes a skill (resource `studio://skill`, prompt `studio`) with every tool, the model's writing rules and step-by-step recipes. Save `docs/mcp-skill.md` as `~/.claude/skills/yue2-studio/SKILL.md` to keep it in Claude Code (see [[AI Agent Skills]]) - The repo ships an `llms.txt` with the same content for tools that look for one (see [[llms.txt convention]]) - In the settings, "Agent (MCP)" routes the studio's own "write" buttons to your connected agent instead of the small built-in assistant The README puts a "For AI agents" section right after the intro. I'd like more desktop tools to ship this combination of [[Model Context Protocol (MCP)]] server, skill and llms.txt: the agent gets the tools and the instructions from the app itself, and nobody has to write a wrapper. Only agents on the same computer can connect, and there is no key, so any local process can drive the studio. ## Sibling studios timoncool built the same studio shell around other models. YuE2 Studio grew out of MiniMax Music3 Studio, and LoRA training datasets move between the two as a folder. ACE-Step Studio runs ACE-Step and is billed as "Suno at home". ## Limits and licenses - Windows 10/11 x64 only for now. yue2.cpp builds on Linux and macOS (Metal), but the studio's release pipeline only ships Windows builds - The app is MIT; the models are NOT. The YuE2-3B and VAE weights are CC BY-NC 4.0 with an individual-creator permission (2026-09-16): individuals may publish and sell the songs they make, companies need a license. SheetSage2 (used for covers), the decoder companion and the MuScriptor weights are plain CC BY-NC 4.0. The visualiser's spectrum component is AGPL-3.0 - If SheetSage2 gets the melody wrong, the cover sings it wrong - Bus factor of one: timoncool has 260 commits, the next human contributor has 4 (the rest come from a GitHub Actions bot) - Three major versions in nine days, so expect breaking changes ## References - GitHub repository: https://github.com/timoncool/YuE2-Studio - Project page: https://timoncool.github.io/YuE2-Studio/ - yue2.cpp: https://github.com/ServeurpersoCom/yue2.cpp - GGUF weights: https://huggingface.co/Serveurperso/YuE2-GGUF - YuE2 model license: https://github.com/multimodal-art-projection/YuE/blob/main/MODEL_LICENSE - README and `llms.txt` (fetched 2026-10-04): features, hardware, model sets, MCP server, licenses - GitHub API (fetched 2026-10-04): repository metadata, releases, contributors ## Related - [[YuE]] - [[Model Context Protocol (MCP)]] - [[llms.txt convention]] - [[AI Agent Skills]] - [[Claude Code]] - [[Rust]] - [[React]] - [[Tauri]] - [[llama.cpp]] - [[Whisper]] - [[Gemma]] - [[OpenRouter]] - [[Local-First Software]] - [[Open Source]] - [[Running AI Models Locally]]