Make full songs from a style description and lyrics on your own hardware: a phone, a foldable, a Mac or a Linux PC. No server, no account, no upload; after the model download it works offline.
The app is model-agnostic. Models come from a registry (assets/models/registry.json) that
declares each model's files, licence, features and limits. The first model is
YuE2-3B by M-A-P (48 kHz stereo songs with vocals).
What you can do: generate one or several variants, queue work and walk away, play and export, re-render with another seed or step count, extend a song from any section (chained extends), reuse a song's prompt for a new take, and rename songs.
Website, with a short ad and a song the app made: https://cypherpunk.today/studio-localhost/
The engine is chosen by hardware (you can override it in the model setup):
| Hardware | Engine | Notes |
|---|---|---|
| Apple Silicon Mac | mlx-Yue (MLX) in a Python sidecar | Measured on an M2 Max: a 25 s song in ~23 s at 8 steps (8-bit AR); extend +15 s in ~14 s |
| Linux, Intel Mac | yue2.cpp (ggml: CUDA, Vulkan, Metal or CPU) via the same sidecar | built locally by the setup flow |
| Android (arm64, Android 14+) | yue2.cpp in-process (JNI, CPU) in a foreground service | jobs run in the background and resume after being killed |
The app is free software under the MIT licence. Each model has its own licence, shown in the model setup and accepted before anything is downloaded; every song records the model and licence it was made with. YuE2-3B weights are CC BY-NC 4.0 (non-commercial use, attribution to M-A-P). Third-party code is listed in THIRD_PARTY.md.
Download sizes: about 3.5 GB on Android (Q6_K GGUF + VAE), about 10.5 GB on Apple Silicon (MLX; mlx-Yue checks both AR precisions, so both are downloaded), 3.2–4.3 GB for GGUF on Linux.
- Android: Zapstore, or the signed APK from the releases page. arm64-v8a, Android 14+.
- macOS:
Localhost Recording Studio.app(Apple Silicon recommended). Not notarized yet: right-click > Open the first time. - Linux: build from source (below).
On first start the app opens the model setup: pick the weights, read and accept the licence, and it downloads them (and on desktop installs the engine runtime) for you.
Requirements: Flutter 3.47 (Dart 3.13), git with submodules.
git clone --recurse-submodules https://github.com/jooray/studio-localhost
cd studio-localhost
flutter pub get
flutter testmacOS: Xcode, then flutter build macos --release.
At runtime the setup flow needs git (Xcode command line tools). It uses
uv if you have it; otherwise it fetches the pinned standalone uv
binary (sha256 checked) into the app's data folder, ~/Library/Application Support/Localhost Recording Studio/bin. Nothing is installed system-wide. The weights land in
models/yue2-3b/mlx/ as mlx-Yue2-3B/ (exactly the converted file set) plus vae/ beside it,
and setup runs mlx-Yue's verify_conversion before the first song. A folder left by 0.4.4 is
repaired in place. An existing mlx-Yue checkout can be used via
"Use an existing mlx-Yue folder" in the setup, or STUDIO_MLX_ENGINE_DIR=/path/to/mlx-Yue.
Linux: flutter build linux --release. Build dependencies (Debian/Ubuntu names):
clang cmake ninja-build pkg-config libgtk-3-dev libmpv-dev. At runtime: git, cmake and a C++
compiler (the setup builds yue2.cpp; CUDA or Vulkan headers are used when present). uv is fetched
into $XDG_DATA_HOME/localhost-recording-studio/bin if it is not installed.
Android: Android SDK with NDK 28.2.13676358 and CMake 3.22.1, JDK 17 or newer (Android
Studio's JBR works). flutter build apk --release --target-platform android-arm64.
The yue2.cpp submodule is patched at build time (android/patches/yue2-android-hooks.patch).
Release signing reads ~/.apk-signing-keystore/signing.properties (or
STUDIO_SIGNING_PROPERTIES) with storeFile, storePassword, keyAlias, keyPassword;
without it the release build is signed with the debug key.
Sidecar (desktop engine) by hand:
cd sidecar
uv sync --frozen --no-dev # numpy + soundfile, for the yue2.cpp backend
# MLX: run the sidecar with mlx-Yue's own venv instead
STUDIO_MLX_ENGINE_DIR=/path/to/mlx-Yue python tools/e2e.py /path/to/mlx-Yue/.venv/bin/python /tmp/e2elib/
core/ no widgets: model registry, engines, library, job queue, setup, settings
models/ registry.dart (reads assets/models/registry.json)
engine/ Engine interface; SidecarEngine (desktop), AndroidNativeEngine (channels);
engines.dart picks the engine from Hardware
library/ song.json format, Library (shared by every platform)
jobs/ JobQueue (persistent queue = songs with status "queued")
setup/ ModelSetup: the one first-run flow (licence, sizes, engine, download)
state/ controllers the screens bind to (forms, navigation, library actions)
ui/ widgets only: adaptive shell (compact / medium / expanded by window width),
screens, theme
sidecar/studio_engine/ Python sidecar (JSON lines on stdio); yue2/ holds the YuE2 backends
android/app/src/main/
kotlin/…/ MainActivity (channels), jobs/GenerationService (foreground service, queue,
resume), models/DownloadService (resumable, SHA-256 checked), yue2/ (JNI adapter)
cpp/ yue2_jni.cpp, CMakeLists.txt, yue2.cpp submodule
Every song is a folder with song.json (title, parameters, model and licence, status, timings),
the audio, tokens.csv (the model's token stream for the whole timeline), score.abc and a
work/ folder with engine checkpoints. Re-render, resume and extend read only this layout, so
they behave the same on every platform.
Below the library, cache/covers/v1-<sha256>-<mode>.abc keeps the score heard in a recording.
It is keyed by the recording's bytes, so a later cover of the same file (another take, another
session, a deleted earlier take) skips listening and goes straight to generation.
Self-test (desktop): STUDIO_SELFTEST=<report dir> STUDIO_SELFTEST_QUIT=1 runs a short
generate at 8 steps and an extend through the real app path and writes selftest.json.
- YuE2 model: M-A-P (Multimodal Art Projection)
- yue2.cpp and the GGUF conversions: ServeurpersoCom
- mlx-Yue and the MLX conversions: vanch007
- ggml, MLX, media_kit, Flutter
Localhost Recording Studio is on Nostr as studio-localhost@cypherpunk.today:
npub10ykef0updslee7dy0hz9rcy6ly3ptt25hcuhe4cank7lk276k5fsrfqdhf