MiniMax Music3 Studio
GitHubGitHub stars Download

Full-length AI music on your own GPU

A Windows desktop studio for MiniMax Music3.

Download for Windowsv3.8.5 · 432 MB Star on GitHubGitHub stars

Windows 10/11 x64 and an NVIDIA card, from the GTX 900 series on. The model files are downloaded from inside the studio, when you decide to — nothing downloads on its own.

Writing a track: the caption as a document with sections, the lyrics, and every generation parameter on a slider.

What it does

Your own LoRA

Your own LoRA trained on the card: the loss as it learns, a checkpoint every 100 steps, one click into the library.

Songs of one artist or style become a LoRA on your card, with HOT-Step’s trainer and recipe.

Word-level karaoke

A finished track: cover, timed lyrics, the request that produced it, and the player.

Enhanced LRC with per-word timings, aligned by Parakeet, Whisper or a cloud model — your lyrics, only the timing is borrowed.

Stems on your GPU

Studio tools: six-stem separation on the GPU, transcription and the wave editor.

Six tracks - drums, bass, other, vocals, guitar, piano - separated by HT-Demucs on the card, or on the processor when you prefer.

Any track to MIDI

Any track to MIDI: a piano roll of every instrument, played against the original, with mute and solo.

A song, a stem or a processed take becomes multi-instrument MIDI - 34 instrument groups and drums - with MuScriptor on your GPU.

A cover for every track

A cover for every track: Commons photos for the style, patterns, generation or your own file.

A track without one wears a free CC0 photograph from Wikimedia Commons that fits its style, a pattern in one of 21 styles, or a cover generated through OpenRouter.

Driven by your agent (MCP)

Connect an agent over MCP: the address and the lines to paste into Claude Code or any other client.

Connect Claude Code, Claude Desktop or Cursor and the agent does everything the studio does: writes and makes songs, fills a dataset and trains, draws covers, cuts clips in the video editor, plays songs - and sees the window and presses its buttons.

Community reviews

Video reviews and tutorials made by the community.

Made a review or tutorial? Put a link to the project in the description — we feature the best here. Send yours through Issues →

Screenshots

The studio itself, in your language.

Model sets: one quantisation per role, what is already on disk, and switching between installed sets.
Model sets: one quantisation per role, what is already on disk, and switching between installed sets.
Every capability runs where you say - locally or through OpenRouter, with the model named.
Every capability runs where you say - locally or through OpenRouter, with the model named.
Cover art: a large preview, prompt templates filled in from the track, or a file of your own.
Cover art: a large preview, prompt templates filled in from the track, or a file of your own.
The LoRA catalogue: styles, artists and sound, each credited to its author, downloaded in one pass.
The LoRA catalogue: styles, artists and sound, each credited to its author, downloaded in one pass.
A dataset: the lyrics found for every song, the caption written by ear with the tempo and key measured.
A dataset: the lyrics found for every song, the caption written by ear with the tempo and key measured.
Audio processing: noise reduction, the Spectral Lifter, a vocal naturaliser, VST3 plugins and mastering to a reference.
Audio processing: noise reduction, the Spectral Lifter, a vocal naturaliser, VST3 plugins and mastering to a reference.
Tracks made by tools stay in the library, each linked to the one it was made from.
Tracks made by tools stay in the library, each linked to the one it was made from.
The equalizer with its curve and MilkDrop over the studio, as the song plays.
The equalizer with its curve and MilkDrop over the studio, as the song plays.
The whole window as Winamp 2: equalizer, playlist and MilkDrop, skinned.
The whole window as Winamp 2: equalizer, playlist and MilkDrop, skinned.
The MIDI editor: tracks and notes with editable chords and section markers.
The MIDI editor: tracks and notes with editable chords and section markers.
The prompt checked as you write: the length by the engine's own tokenizer, words after a section tag that would not be sung, and a caption that does not say who sings.
The prompt checked as you write: the length by the engine's own tokenizer, words after a section tag that would not be sung, and a caption that does not say who sings.

Getting started

  1. 1InstallRun the installer or unpack the portable archive.
  2. 2Choose a setThe first screen preselects what your card can run; press download and it fetches only what is missing.
  3. 3WriteA caption is the style, the lyrics are optional — an instrumental is one switch.
  4. 4GenerateThe engine starts by itself and the track lands in your library, with its settings attached.
  5. 5Optional: connect OpenRouterOne key adds the writing assistant, cover art and cloud transcription. Local stays the default.
Download for Windowsv3.8.5 · 432 MB

Questions

What graphics card do I need?

Windows 10/11 x64 and an NVIDIA card, from the GTX 900 series on. The model files are downloaded from inside the studio, when you decide to — nothing downloads on its own.

A runnable Music3 installation is always five components: language model, RVQ depth decoder, condition encoder, DiT and vocoder.

Your GPUProfileDownload
30 GB+Full Native — BF16 / BF16 / F3228.6 GB
15 GB+Q8 Quality12.8 GB
11.5 GB+Balanced — Q6 / Q8 / Q59.8 GB
9.5 GB+Light — for speed and low VRAM7.7 GB
8 GBMinimal — Q3, for 8 GB cards6.5 GB

The studio detects your card and preselects the profile it can actually run, but the download is always your decision. Every file is checksum-verified against a pinned Hugging Face revision, and an interrupted download resumes where it stopped.

Is it free?

Yes. The studio is free and open source, and the models download for free from inside it.

Does it work offline? Is my music sent anywhere?

The short version: your music is yours, and it stays on your disk.

  • No account — There is nothing to sign up for and nothing to sign in to.
  • No telemetry — The studio does not report what you generate, or that you generated anything.
  • Cloud is opt-in — Nothing leaves the machine until you add a key, and then only for the capability you assigned it to.
Can I use the songs I make?

The studio is open source. MiniMax Music3 weights are governed by their own community license — commercial use must display the MiniMax-Music3 name and implement the safeguards that license requires.

Can I train it on my own style?

Songs of one artist or style become a LoRA on your card, with HOT-Step’s trainer and recipe. Every setting is editable, and datasets move between this studio and YuE2 Studio.

Can an AI agent use it?

Connect Claude Code, Claude Desktop or Cursor and the agent does everything the studio does: writes and makes songs, fills a dataset and trains, draws covers, cuts clips in the video editor, plays songs - and sees the window and presses its buttons. MiniMax’s own caption rules come with it, so the agent writes the captions and lyrics itself. Pick it as the writing assistant and it writes instead of the local model.

What is inside the download?

Local by default. A cloud key is added by you, for the capability you choose, and can be taken away again.

PartRunsSize
MiniMax Music3 engine (minimaxmusic.cpp)Your GPU, CUDAbundled
Music3 model set - five componentsYour GPU6.5-28.6 GB
Karaoke timings - Parakeet or WhisperYour machineoptional download
Stem separation - HT-Demucs, six stemsYour GPU or CPU136 MB
Writing assistantLocal GGUF or OpenRouteryour choice
Cover artOpenRouter image modelcloud only
Video export - ffmpegYour machinebundled
LoRA trainer — HOT-Step ace-trainNVIDIA RTX 30 or newer, 22 GB of VRAM10.5 GB, optional
Describing songs by ear — MOSS-Music-8Babout 12 GB of VRAM10 GB, optional
VST3 host — HOT-StepYour machine, a process of its ownbundled
Audio to MIDI — MuScriptor (HOT-Step ace-midi)NVIDIA GTX 16 / RTX 20 or newer0.5–5.6 GB, optional
How is it built?

The service is compiled into the desktop binary and supervises the C++ engine. When the application closes, everything it started closes with it.

React UI ─┐
          ├─ MiniMax Music3 Studio.exe   (window + native service)
Rust axum ┘        │
                   └─ minimaxmusic.cpp `mm-server`  (C++/CUDA, GGUF)

Who made this

Nerual Dreming — artist, founder of ArtGeneration.me and of the Neuro-Cartel community, author of ACE-Step Studio, which this studio grew out of.

GitHub · @timoncool Telegram · @nerual_dreming Telegram · @neuroport ArtGeneration.me Support the project