Songs from words and chords
on your own machine

Seven rooms, one studio

The Creator makes the song, the Writer writes it with a chat model, the Refiner finishes it, the Artist paints its pictures, the Videographer plans its music video, the Librarian keeps every take, and the LoRA Trainer teaches YuE2 your own songs. A guide for every room is inside the Studio.

Creator

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

Writer

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

Refiner

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

Artist

Pictures for your songs from a prompt of your own: square, wide and tall from one seed, painted by Qwen-Image 2512; the picture develops before your eyes, and a square one becomes the take's cover. An early preview.

Videographer

A music video for the take: its song cut into shots at the sung lines, the scenes written with the Writer, the key frames painted by the Artist, the moving shots by MiniMax H3; a timeline to edit, a look (3D LUTs, grain), an animatic and the clip. Beta.

Librarian

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

LoRA Trainer

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

The cheat-sheet ♪

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

The guide · F1

The whole Studio explained inside the Studio, a key away: every room and every knob, what it does and the numbers we measured, in twelve languages. Made for the joy of making, not for hours of searching the web.

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 135 s; a 7:25 song with its score first in 212 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.
• Lyrics that keep time. Each line's beats against its section's ruler: syllables by each language's own rules, and the consonants that take a beat of their own; the section tags offered as you type «[»; stress marks astray found.
• 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, and its guide, in English, Russian, Ukrainian, Belarusian, Greek, Spanish, Italian, French, Portuguese, German, Japanese and Chinese.

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.

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.

Lyrics that keep time

Beside each line its beats, against the ruler of its section: verse and bridge one group, the pre-chorus and the chorus their own. A syllable is counted by each language's own rules, and the consonants that take a beat of their own count too; a pause or an interlude inside a section opens no new group.

Type «[» at a line's start and the section tags are offered, as a code editor offers its words. Stress marks astray are found; Ctrl+F finds in the box alone.

Ruach Studio and SUNO, honestly

Every week a video says some model kills SUNO. This is not a model: it is a whole studio around one, with a fine-tune of its own. Where it wins, and where SUNO still does, said plainly.

«—»: not that we found, in more than half a year on SUNO's paid plans. SUNO moves fast; this is November 2026.

The ring: YuE2 against SUNO v6 and Musicful

Twenty cards, each a style line and lyrics, sent unchanged to every side. SUNO v6 and Musicful V3 as their services deliver them; ours raw from the engine, sung again in November with every peak whole: the stock YuE2-3B, our own fine-tune RS-FW at three of its saves, and RS-TW 2.0. Here every file plays at the same loudness, so none wins by being louder. Press a name to listen; another name on the same card goes on from the same second.

RS-TW 2.0 is our second fine-tune: its epoch 2.0, merged over RS-FW 2.4 at two strengths, ×0.75, the safe balance, and ×0.9, more tectonic, with no promise of stability. Ours are two seeds a card, both here, neither chosen over the other: the switch in the player gives seed 1 or seed 2 and keeps it from card to card, and the table's means count both.

first · second · third of the eight in every row, the first also bold; ↥ higher is better, ↧ lower is better, → the closest to the value shown, the true peak from below. Cells that show the same number share a colour.

Every side is heard on an RTX 3090, the mix and the voice alike, by Whisper large-v3: the first Whisper row on the mix, the second on the separated voice, where BS-Roformer splits all of a take's voices from the music, the lead, the backing voices and the choir together, never the lead alone. Without the music Whisper sometimes loses whole lines and writes words nobody sang into the pauses, on every side, so the two rows are best read together.

FW and TW are the code names of our two fine-tunes of YuE2 3B: the First Wave and the Tsunami Wave. The Tsunami Wave was trained on 109,250 tracks, about 6,650 hours: 66.7% of the hours instrumental, 31.6% sung, 1.8% spoken. The voices are in English (7.0% of the hours), Russian (5.8%), Hebrew (1.7%), Greek and Ukrainian (0.6% each), French and Spanish (0.5% each), Portuguese (0.3%), Italian (0.2%) and fourteen more languages under 0.1% each; in 15.9% of the hours a voice is heard whose language was not identified.

Theirs come mastered, ours raw at the decoder's own level with every peak whole: loudness is left to the master, which is the Studio's Refiner's to do. Other seeds and other prompts sound different on every side; the ring shows the quality and the spread, not a final ranking. The whole table and how it was measured: the model card.

What the spectrum says

The ring measured instead of heard: 280 takes, 40 for each of our six sides, 20 of SUNO v6 and 20 of Musicful V3. Every take was first brought to −14 LUFS, so loudness wins nothing, and every curve is its take's own level, band by band, against that take's whole power: only the shape is compared. Ours are the engine's raw output and theirs come mastered, so some of the gaps below are the master's.

Measured with our own code on the CPU: 1/3-octave bands from a Welch spectrum, loudness and its range by ITU-R BS.1770-4 and EBU Tech 3342. Each picture opens larger on a click.

What the stereo says

The ring's 280 takes, measured for their width. In each of four bands the side, half the difference of the two channels, is set against the mid, half their sum: the higher the side over the mid in dB, the wider that band sounds, and a take in mono would have no side at all. The panorama is how far the difference in level between the left and the right channel swings over a take, from its 5th to its 95th percentile, in dB. Each number is the median of a side's takes, 20 for SUNO v6 and for Musicful V3, 40 for each of ours. Wider is not better here, only wider: this table has no places.

Our merges keep the mids and the highs where the stock YuE2-3B has them, none more than 1.3 dB away. RS-TW 2.0 is a little narrower at the edges: ×0.75 by 1.6 dB in the bass, 1.0 dB in the air and 0.4 dB in the panorama, ×0.9 by 2.7, 1.4 and 0.5 dB. SUNO v6 and Musicful V3 open the bass 4.9 and 4.4 dB wider than the stock YuE2-3B.

To Viktor's ear the narrower edges of RS-TW 2.0 sound steadier, not like lost stereo: the stock YuE2-3B often parks a guitar in one channel for 15 to 20 seconds.

Measured with our own code on the CPU: a short-time spectrum of 4096 points every 50 ms over the first 240 seconds of each take, the mid and the side as (L+R)/2 and (L−R)/2, the panorama from the difference in level between the channels in 200 Hz – 8 kHz, in the moments within 40 dB of the take's loudest.

What you need

What it downloads

Measured file by file against the pinned lists of the Studio's repositories and on a fresh install's disk (1 GB = 10⁹ bytes). The install brings the Creator, the Librarian and the Refiner's chain; every other room comes only when you ask, from Engine → Models and LoRAs → By room, which measures the free disk first and fetches nothing that would not fit.

Only when you ask for it by name: the other sound engines (the IRA, IRAE and RS-FW merges, 35 to 86 GB by the quant), the Artist's adapters (7.5 GB) and fine-tunes (20.4 GB each), MiniMax H3's every tier (201 GB) or BF16 (478 GB), the 💎 workspaces (5.6 GB). Not counted: the NVIDIA driver and the CUDA toolkit, Pianoteq (yours to buy), your takes.

What comes next

2.0.5 is the Studio we make our own songs in every day. These are the larger pieces on their way, for 2.1.0 and after.

• Songs edited as the model writes them. A take extended from any moment; a section written anew with its neighbours' voice, key and bars kept, joined in the model's own tokens; an instrumental layer on the take's own score.
• The DAW, further. A chord lane and a beat grid over the take, a reference drawn from any take, and a piano roll that writes what you play on your MIDI keyboard into the score.
• The Videographer, further. The lyrics sung on the clip, word by word, and one colour across all its scenes.
• The next wave of our fine-tune. Taught on what YuE2 sings worst, and heard blind before it ships.
• Native plugins for REAPER, Waveform and Bitwig, once the Studio has found its people.

Licences

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)
• Qwen-Image 2512 by Qwen, with Lightning by lightx2v: the artwork
• YuE2 Studio: the score checker; GTSinger: voices to train on