Remiqora
  • Open source · Runs locally
  • New: From an idea · Stable Audio 3
  • Windows · NVIDIA CUDA
  • macOS · Apple Silicon

Made with AI.
Made by you.

Remiqora is a local music studio. ACE-Step 1.5, YuE2-3B and Stable Audio 3 generate the track, from your prompt or from one phrase, Demucs splits it into stems, the AI arranger writes new parts that fit, and a multitrack editor in your browser mixes it down. Everything runs on your own computer.

Get it on GitHubWatch the demo

  1. From an idea
  2. Generate
  3. Stems
  4. AI arranger
  5. Mix
AI arranger in the DAW
The editor with four stem lanes and the AI arranger panel: Add a part, the instrument chips with Strings picked, a part description and the lanes the model hears
Effects rack
The effects rack of the drums channel with EQ, dynamics, filter, chorus, delay, reverb and distortion switched on
Stem separation
The stems panel: vocals, drums, bass and other, with Recreate, Open in editor and Delete

Listen to a song come together

Everything here was made in Remiqora for this page. ACE-Step 1.5 wrote the song from one prompt, Demucs split it into four stems, and the AI arranger wrote three string parts over them. Press play, mute or solo the stems, and switch the strings on while it plays.

0:00 / 0:42
0:000:100:200:300:40
Vocals
Drums
Bass
Other
StringsAI arranger
Strings from the AI arranger
The song
energetic pop rock, powerful female vocal, punchy live drums, driving bass, crunchy electric guitars, anthemic chorus, 128 bpm, no strings

ACE-Step 1.5 turbo made three 1:40 variants in about 15 seconds once the model was loaded. One of them plays here, untouched.

The strings
driving staccato string section, rhythmic violins and cellos, energetic, building into the chorus

The AI arranger heard the whole mix and wrote three takes in about 40 seconds on the base model. They play 6 dB down, as you would pull the fader.

Times on an RTX 4080 with 16 GB. The lyrics were written for this demo; the Russian page has its own song with the same prompt.

Four prompts, four songs

The first variant of each prompt as ACE-Step 1.5 turbo gave it: no editing, no mixing. Each run made two one-minute variants in about 10 seconds on an RTX 4080. Here are 30 seconds of each.

  1. deep house, rolling bassline, warm analog pads, crisp hi-hats, four on the floor, 122 bpm

    Instrumental, 0:30

  2. lo-fi hip hop, dusty boom bap drums, mellow Rhodes chords, vinyl crackle, upright bass, 80 bpm

    Instrumental, 0:30

  3. cinematic orchestral, soaring strings, french horns, taiko and timpani, epic build

    Instrumental, 0:30

  4. russian indie rock, male vocal, jangly electric guitars, driving drums, bass, energetic

    With Russian lyrics, 0:30

From prompt to mixdown

  1. Start from an idea

    Write one phrase: what the song is about, the mood, who it is for. A local language model writes the prompt in the chosen engine's own format, with style tags, lyrics, tempo and key, and you can edit it before generating. Ask for rain on a window and Stable Audio 3 makes a sound, not a song.

    From an idea: a one-line idea, the engine picker and the prompt the assistant wrote, with style, lyrics, tempo and key
    The prompt the assistant wrote from one phrase

  2. Generate

    Describe the music, or bring a style and lyrics. ACE-Step 1.5, YuE2-3B and Stable Audio 3 share one interface: pick a model in the header and Remiqora starts it and stops the other, so a single GPU is enough.

    The generation form with a preset picker, a description field, duration and number of variants
    The generation form and your track library

  3. Split into stems

    One click separates any track into vocals, drums, bass and the rest with Demucs. Open in editor sends all four into a new project.

    The stems panel: vocals, drums, bass and other, with Recreate, Open in editor and Delete
    Four stems, each with its own player and download

  4. Add parts with the AI arranger

    Pick an instrument and the lanes the model should hear: it writes drums, bass, strings or vocals in their tempo and key. It can also repaint a section, continue the track or cover it in another style. Every run gives 2 to 4 variants, each on its own lane, so you keep the best one.

    The editor with four stem lanes and the AI arranger panel: Add a part, the instrument chips with Strings picked, a part description and the lanes the model hears
    The AI arranger docked next to a four-stem project

  5. Mix in the editor

    Trim, split and fade clips, loop a section, snap to a BPM grid, and time-stretch a clip to the project tempo without changing its pitch.

    The editor timeline: four colored stem lanes with their volume sliders, the drums lane selected, the master meters and the panel buttons
    The timeline with four stem lanes, the drums lane selected

  6. Export

    Render the mix as WAV or MP3 in the browser and it is saved back to your library, or export every lane with a Reaper project or a DAWproject file to finish in your own DAW.

A full effects rack on every channel

Every track and the master bus carry the same eight effects, all running in real time: EQ, compressor, filter, chorus, delay, reverb, distortion and bitcrusher.

Also in the box

  • Stable Audio 3 for instrumentals, sound effects and samples in seconds (Powered by Stability AI)
  • LoRA training for ACE-Step, right in the interface
  • Melody extraction with SheetSage2 and audio-to-MIDI with MuScriptor
  • Undo and redo for 30 steps, and a detailed built-in guide for every section
  • Interface in 20 languages, in your system language by default
The effects rack of the drums channel with EQ, dynamics, filter, chorus, delay, reverb and distortion switched on
The channel strip of a drum track

Speaks your language

The interface, the built-in guides and the installer come in 20 languages, Chinese in both scripts, and the idea for a song can be written in any of them. Remiqora opens in your system language; the switcher in the header changes it.

Describe the song in your own language

  • Write the idea in Spanish, Arabic, Hindi or Kazakh: Remiqora recognises the language from the letters and the words, and the local assistant writes the lyrics in it.
  • The language of the song can also be picked by hand from 51 vocal languages.
  • Arabic, Persian and Hebrew are mirrored right to left. The timeline, the waveform and the mixer stay left to right, as time and signal flow do.

Tried in every language

  • We wrote an idea in each language and checked what the assistant answered: the lyrics came back in that language every time.
  • We generated one short song per language in ACE-Step and in YuE2 and ran them through Whisper speech recognition. YuE2 sang about 90 % of the words on average and ACE-Step a bit over half; both vary a lot from language to language. It is a sample, not a promise.
  • Kazakh works too: ACE-Step does not list it, but it sings it, with an accent.

Every language except Russian and English is machine-translated and has not been read by a native speaker, so some phrases will sound odd. Found one? Tell us on GitHub.

Or use it as a desktop app

On Windows and on Macs with Apple Silicon, Remiqora also comes as an installer. It opens in its own window and sets everything up on the first launch: no Git, no Python, no compiler.

Download for WindowsDownload for macOS

Version 0.5.0, pre-release. Windows: NVIDIA GPU. macOS: Apple Silicon. See all files and SHA-256 sums.

  • It checks your GPU, driver and free space and asks what to install: ACE-Step always (about 18 GB), YuE2, stem separation, Stable Audio 3, the prompt assistant and the AI arranger models as options, each with its size.
  • Anything left out shows up as not installed, with a button that adds it later while the app keeps running.
  • Closing the window stops the model servers and frees the GPU.
  • Models, projects and logs stay in that folder. Nothing is uploaded anywhere.

Built with Electron around the same web interface. The first launch uses uv for Python, the prebuilt audio.cpp engine (CUDA or Metal) and FFmpeg.

Experimental. The installers are not signed yet, so Windows shows a SmartScreen warning and macOS may ask you to allow the app. Prefer to build it yourself? The steps are in the desktop folder.

How the desktop app works

The first-launch screen: graphics card, driver, free space and connection checked, the parts to install with their sizes, and a folder chosen for models and projects
First launch: the app checks your computer and asks what to install
Components downloading and installing, with overall progress and progress for each component
Then it downloads and installs the components, with progress

What to know before you install

Works today

  • Windows with an NVIDIA CUDA GPU. Tested on an RTX 4080 with 16 GB.
  • macOS on Apple Silicon. Tested on an M5 with 24 GB; newer and slower than the Windows path.
  • Linux x86_64 with an NVIDIA GPU, via community-contributed scripts.
  • Stem separation, MIDI transcription and the editor.

Limits

  • No desktop installer for Linux yet — run from source via the scripts.
  • The engines run one at a time by default.
  • Output quality is whatever the models give you.
  • The desktop installers are experimental and unsigned.
  • In active development: releases are pre-releases for now.
  • Licenses: Remiqora and ACE-Step 1.5 are MIT, but the YuE2-3B weights are CC BY-NC 4.0, so no commercial use, and Stable Audio 3 is free for commercial use only below $1M of annual revenue.

Install

With the scripts you need Git and they install the rest; every option and troubleshooting step is in the README. Prefer no terminal? Use the desktop app above.

Windows with an NVIDIA GPU

git clone https://github.com/inikolax/remiqora.git
cd remiqora
setup_prereqs.bat
setup_models.bat
dev.bat

Install the NVIDIA driver yourself: the script leaves it alone. Model downloads are about 10 GB.

macOS with Apple Silicon

git clone https://github.com/inikolax/remiqora.git
cd remiqora
./setup_prereqs.sh
./setup_models.sh
./dev.sh

Uses a prebuilt Metal build of YuE2, so no compiler is needed.

Linux x86_64 with an NVIDIA GPU

git clone https://github.com/inikolax/remiqora.git
cd remiqora
./setup_linux.sh
./prod_run_linux.sh

Community-contributed. Builds audio.cpp from source, so you'll need a CUDA toolkit (nvcc, cuBLAS, cuFFT) besides the driver. Verified on Ubuntu 24.04.

Read the full installation guide

Questions people ask

Is Remiqora free?

Yes. It is open source under the MIT license, with no account, no subscription and no limit on how many tracks you generate. It runs on your own hardware, so the only cost is your electricity.

Does it work offline?

After the one-time download of the models, yes: generation, stem separation, the AI arranger and the editor all run on your computer without an internet connection. Remiqora has no telemetry of its own, and your prompts and tracks never leave your machine.

What hardware do I need?

Windows or Linux with an NVIDIA GPU, or a Mac with Apple Silicon. It is tested on an RTX 4080 with 16 GB and on an M5 Mac with 24 GB. YuE2 has a q4_0 build for cards with 8 GB, its q8_0 build wants about 9 GB, and the optional XL base model needs 16 GB. Plan for 18 GB of disk for ACE-Step alone and up to about 80 GB for everything, Stable Audio 3 and the prompt assistant included.

Can I use the music commercially?

It depends on the model that made the track. ACE-Step 1.5 is MIT-licensed, and its authors say the music can be used commercially. The YuE2-3B weights are CC BY-NC 4.0, so tracks made with YuE2 are for non-commercial use only. Stable Audio 3 weights are under the Stability AI Community License: commercial use, of the generated audio too, is free for individuals and companies with under $1M in annual revenue after a free registration with Stability AI; above that you need a license from Stability AI. Check the model's current license before you publish, and note that the ACE-Step authors warn about unintentional copyright infringement.

How is it different from Suno or Udio?

Suno and Udio are cloud services with their own models. Remiqora runs open models on your computer: no credits, no queue, and nothing is uploaded. Instead of one finished file you get stems, a multitrack editor and an AI arranger, so you can build a song part by part and finish it in your own DAW. The cloud services can still sound more polished in some genres.

Does it run on Mac and Linux?

Macs with Apple Silicon get a desktop installer; Intel Macs are not supported. Linux with an NVIDIA GPU runs from source with the setup scripts; there is no Linux installer yet.

Something else? Ask in the GitHub issues.