Kwesi 0.2.1

What you make is yours

• The license now says so. Kwesi's code stays under CC BY-NC 4.0, but Fobia adds a permission on top: the music, audio, MIDI, notation, lyrics and trained models you create with Kwesi are yours. Fobia places no restriction of its own on how you use them, including commercially.
• Each model's own license still applies. ACE-Step 1.5 (MIT) allows commercial use. MusicGen, YuE2 and RAVE weights are non-commercial. See LICENSES/LICENSE-Kwesi.md for the full terms.
• Training terms, asked once. Before your first training run, Kwesi asks you to confirm that you own or have the rights to everything you train on.

Also in this release

• Updating from 0.2.0 works from inside the app: Settings → About → Download, then Restart to update.
• ESLint is set up for development, and npm run lint works again.

Downloads

Platform	File
macOS, Apple silicon (M1 and later)	Kwesi-0.2.1-arm64.dmg
macOS, Intel	Kwesi-0.2.1-x64.dmg
Windows (64-bit)	Kwesi-Setup-0.2.1.exe
Linux	Kwesi-0.2.1.AppImage or kwesi_0.2.1_amd64.deb

The .zip, .blockmap and latest*.yml files are for the in-app updater. You don't need to download them.

macOS: this build isn't signed or notarized by Apple yet. The first time you open it, macOS may block it. Go to System Settings → Privacy & Security and click Open Anyway. If macOS says Kwesi "is damaged and can't be opened", run this once in Terminal after moving the app to Applications:

xattr -dr com.apple.quarantine /Applications/Kwesi.app

Intel Macs: the app runs, but most models need PyTorch versions that are no longer built for Intel Macs, so their environments may not install.

Windows: SmartScreen may warn about an unrecognised app. Click More info → Run anyway.

Models, their Python environments and weights are installed from inside the app (Model Manager and Settings → Environment). They aren't bundled.

Full Changelog: v0.2.0...v0.2.1

Kwesi 0.2.0

Train your own models

• Training works end to end for all four trainable models. ACE-Step 1.5, MusicGen, RAVE and MuseCoco runs are live in the app, and every trained model can be used for generation.
• Train MuseCoco from your own MIDI files. Instruments, tempo, key and time signature are read from the files, so no labels are needed.
• GPU acceleration for MuseCoco, built once from its card on the Training screen (Linux + NVIDIA). Training then takes about 1 s per update instead of about 50 s.
• Presets with time estimates (Quick test / Standard / Thorough) for each model.
• Checks before a run starts:
  • the dataset (file count, total length, unreadable or empty files, duplicate names)
  • free disk space
• Preview clip when a run finishes, so you can hear the new model right away.
• Continue training a model you already made. Settings that must match the original are locked.
• LoRA strength control when generating with an ACE-Step trained model.
• My Trained Models in Model Manager, with Preview, Open folder and Delete for each model.
• Idle model servers are stopped to free the GPU for a training run.

Updates inside the app

• Settings → About shows your version and lets you check for updates, download one with a progress bar, and restart to install it. If anything fails, you can retry.
• A small update available button appears in the side rail when a newer version is out.
• Nothing downloads until you ask. There's an on/off switch for update checks.
• This is the first release that can update itself. Later versions will appear in the app.

Look and feel

• A new white backdrop: a dotted grid framed by hand-drawn music doodles, with the app's glass panels floating on top.
• Light theme only. The dark palette has been removed.
• The shared UI components are now used across every screen.

Fixes

• MusicGen training runs no longer fail when started in the app. Leftover training work folders are cleaned up.
• Trained models show by their run name and stay registered after a restart.
• The environment install log no longer shows raw colour codes.
• The New Workspace dialog no longer shows through the model setup dialog.
• The required * on form fields stays on the same line as its label.

Downloads

Platform	File
macOS, Apple silicon (M1 and later)	Kwesi-0.2.0-arm64.dmg
macOS, Intel	Kwesi-0.2.0-x64.dmg
Windows (64-bit)	Kwesi-Setup-0.2.0.exe
Linux	Kwesi-0.2.0.AppImage or kwesi_0.2.0_amd64.deb

The .zip, .blockmap and latest*.yml files are for the in-app updater. You don't need to download them.

macOS: this build isn't signed or notarized by Apple yet. The first time you open it, macOS may block it. Go to System Settings → Privacy & Security and click Open Anyway. If macOS says Kwesi "is damaged and can't be opened", run this once in Terminal after moving the app to Applications:

xattr -dr com.apple.quarantine /Applications/Kwesi.app

Intel Macs: the app runs, but most models need PyTorch versions that are no longer built for Intel Macs, so their environments may not install.

Windows: SmartScreen may warn about an unrecognised app. Click More info → Run anyway.

Models, their Python environments and weights are installed from inside the app (Model Manager and Settings → Environment). They aren't bundled.

Full Changelog: v0.1.4...v0.2.0