Ruach Studio
EN
OSEMSimply awesomeOpen Source, Engaged & Musical

Songs from words, on your own machine.

A whole studio around YuE2, the open song model. Write the style and the lyrics, and the studio composes, sings and renders the song on your own GPU. Your songs, words and settings never leave it.

2.0.0-rc1 · the first release candidate · October 2026
The Creator by day: the song form, the take you are listening to, and every take The Creator by night
8 minwhole songs, up to eight minutes
126 sa 6:12 song on one RTX 3090
5 roomsto make, write, refine, keep and teach
200instruments probed by ear
7languages of the page

Five rooms, one studio

The Creator makes the song, the Writer writes it with a chat model, the Refiner finishes it, the Librarian keeps every take, and the LoRA Trainer teaches YuE2 your own songs. A guide for every room is inside the studio.

Creator

Creator

Style, lyrics, planning mode, decoder, sliders, LoRAs on a road from green to red; the take with its score, seeds and everything to make it again.

Writer

Writer

Songs as documents with versions, text profiles, and a chat model of your own (or OpenRouter) that drafts and revises with you.

Refiner

Refiner

Spectrum, artifacts, debuzz, a lyrics check by Whisper, stems, remaster, upscale; one chain runs them all.

Librarian

Librarian

Every take in workspaces: searched, liked, noted, locked, given artwork, sent to your DAW; a trash that gives back.

LoRA Trainer

LoRA Trainer

Adapters from your own songs: the set, the run, both halves' curves explained, the epochs worth hearing first.

The instruments cheat-sheet

The cheat-sheet ♪

200 instruments probed and judged by ear: what YuE2 really plays, what it plays thinly, what it does not know.

What makes it a studio

Not one more model wrapper: a workstation around the model, open to the bone, every change numbered and documented.

  • Whole songs, up to eight minutes. A 6:12 song in 126 s; a 7:25 song with its score first in 198 s, on one RTX 3090.
  • The score first, and yours to edit. YuE2 writes the melody and chords as ABC before a note sounds: change them, transpose them, bring your own score or MIDI, and the song follows it.
  • Two seeds, not one. Keep the song (the music seed), hear it rendered anew (the sound seed).
  • Your own sound. LoRA training on your own songs, in the studio, unquantized (bf16), with telemetry that tells you which epochs to hear first.
  • Post-production, all local. Stems (BS-Roformer, htdemucs), debuzz, remaster, upscale (UniverSR), a lyrics check by Whisper.
  • Into your DAW (experimental). A REAPER project with the stems, the score as MIDI, the sections and the lyrics on the timeline; DAWproject for Waveform and Bitwig.
  • A writer that knows the engine. The style as short tags, the most important first: the way YuE2 follows best.
  • Several GPUs, several jobs. Synthesis on one card, training on the others; with one card the page tells you what waits.
  • For scripts and agents. An API (OpenAPI 3.1) and an MCP server on your machine and your network.
  • Your language. The page whole in English, Russian, Ukrainian, Belarusian, Greek, Spanish and Italian.
Babel in the score: the run stopped, and why
Babel in the score: caught 15 s into the run

A guard against garbage

When the music half writes a broken score (adapters pushed too far do that), the engine stops the run before any music is built on it, in seconds, and says why: which adapters, how far past their limits, what was written.

Strengths sit on measured roads; stacked adapters share a measured ceiling.

The Trainer's curves
Both halves' curves, explained

Its own LoRA trainer

The studio trains adapters itself, on unquantized bf16 weights. Against Ostris' AI-Toolkit on the same set, the music half's loss agrees epoch by epoch within 0.015.

Rank 16 on 148 songs: 2.1 s a step in 12 GB. Rank 32 on 42 full songs: 6.7 s a step in 18 GB.

Ruach Studio and SUNO, honestly

Every week a video says some model kills SUNO. This is not a model: it is a workstation around one. Here is where it wins, and where it does not (yet).

Ruach Studio (YuE2)SUNO
where it runsyour GPU; your songs never leave ittheir cloud
what it coststhe electricitya subscription and credits
the scorewritten first, shown, editable, yours (ABC, MIDI)hidden
repeat a takeexactly: two seeds you keepno
your own soundtrain a LoRA on your songs, in the studiono training
after the renderstems, debuzz, remaster, upscale, lyrics check, all localstems, in the cloud
into a DAWa REAPER project with the score, sections and lyrics (experimental)audio stems
the codeopen, every change documentedclosed
polish out of the boxless: a mix can buzz (Debuzz helps), lyrics can drift (the lyrics check finds where), some instruments YuE2 plays thinly or not at all (a LoRA teaches them)more
what you needan NVIDIA GPU (24 GB for everything at full precision), Linux, about 120 GB of disk for the models, LoRAs and workspacesa browser

What you need

The machine

  • An NVIDIA GPU: 24 GB for everything at full precision
  • Linux (made on Ubuntu 24.04, CUDA 12.8)
  • About 120 GB of disk: the models (about 100 GB, fetched, pinned and checked by fetch-models.sh), the LoRAs and the workspaces

The screen

  • Made for a browser window of 1920×1080 at 100 %
  • The ideal: 2K (2560×1440), 22″ or larger, 16:9 or wider
  • Two themes for hours of work: Scroll & Brick by day, Scroll by Lamplight by night, and 20 more

Optional

  • A chat server for the Writer (vLLM, LM Studio, Ollama) or an OpenRouter key
  • A DAW: REAPER, Waveform 14+, Bitwig
  • More GPUs: training and the lab's work beside the songs

What comes next

2.0.0-rc1 is the studio we make our own songs in every day. These are the larger pieces on their way, for rc2 and after.

  • The Artist room. Covers in three shapes at once from one seed (1:1 for the album, 16:9 for a video, 9:16 for a short), the title and the artist written on them, the versions of each take's picture.
  • Five more languages. Chinese, French, Portuguese, German and Japanese; the right-to-left ones once the page itself runs right to left.
  • The Writer's models with their prices, from OpenRouter's list, as you type.
  • Calmer rooms. Every button that cannot be undone behind a dialog, the inline notes as tips, one shape for the icon buttons.
  • A voice's kind in the Trainer, measured on speech and on singing.
  • A desktop app. An installable page (PWA) first, then an Electron app that starts and stops its services itself.
  • A score editor as a DAW has one. A piano roll, a chord lane, the lyrics over the notes, a MIDI keyboard to play ideas in.
  • Native plugins for REAPER, Waveform and Bitwig, once the studio has found its people.

Licences

The code

  • The GNU AGPL, version 3 or later: anyone may use, study, change and pass it on, and a fork stays open, with its source.

The name and the art

  • The logo, the name «Ruach Studio», and the audio and pictures published with it: CC BY-NC-ND 4.0.

The model

  • The YuE2 weights are CC BY-NC 4.0. Their licence speaks of the weights, not of the songs made with them: read it before you sell.

Standing on

  • YuE2 and MERT-v2 by m-a-p: the model
  • yue2.cpp by ServeurpersoCom: the C++ engine on ggml, carried here with its own patches
  • YuE2 Kit by IronWolve: the base of the page and the scripts
  • AI-Toolkit by Ostris: the trainer's YuE2 code was ported from it
  • Mothersuperior: the realaudio tokenizer head and the VAE merge
  • Qwen2.5-Omni (the listener), Whisper (the lyrics check), BS-Roformer and htdemucs (stems), UniverSR (upscale)
  • Krea 2 Muse by Stable Yogi and CyberRealistic XL by Cyberdelia: the artwork
  • YuE2 Studio: the score checker; GTSinger: voices to train on

A little story

Viktor Zhuromskyy, the author of Ruach Studio, does not publish it for the hype. He made it for himself. After more than half a year of struggling with Suno on paid plans, and of searching for a model worth working with (ACE-Step, MiniMax and others), he set out to build an audio studio of his own, on his own machine. He settled on YuE2: of all the open models, it is the best value and the most refined.

He did not knock the studio together just to have something that runs. He built it his own way, because his work is his own: his translations, his interlinear texts and much more. He fit into it as much as he could, polished the interface as far as he could, and made working with it not poking and clicking among crutches and potholes, but as productive as it can be.

That is the whole story. Now the studio is yours as well.