YuE Studio 0.4.0: iPhone as a second Neural Engine

iPhone as a second Neural Engine

YuE Studio can now use an iPhone's Neural Engine as a second synthesis engine. With the companion app YuE Remote (app/YuERemote, an Xcode project you build and install with your own Apple developer account) running on a phone on the same network, the Mac finds it automatically, sends the synthesis weights once (2.8 GB), and from then on two queued songs synthesize at the same time: one on the Mac's Neural Engine, one on the phone. A song already running on the GPU hands over to the phone mid-solve when the phone frees up, and anything the phone cannot take goes back to the Mac.

• Measured on an iPhone 17 Pro: the phone runs the same layer program at the M4's pace or better, and its latents match the Mac's to fp16 noise (correlation 0.99999).
• Songs up to about 2 min 40 s go to the phone for now; longer ones stay on the Mac until the larger program compiles on the phone's compiler.
• The phone shows a TFLOP/s tachometer while it works. Keep YuE Remote in the foreground; it keeps the screen awake itself.
• Run panel: "Use an iPhone's Neural Engine" toggle with a status line; the stage track labels songs "iPhone" or "GPU, then iPhone".
• Quality must be a "GPU + Neural Engine" option for the phone to be used.

Also in this release

• Lyric writer: the outro is no longer truncated; writing lyrics no longer carries the theme over from the previous generation.
• App icon, and the macOS app source split into files (thanks to @Shinisaru, PR #3).
• Details of what the phone's compiler accepts are in docs/apple-silicon.md.

Install: open the DMG and drag YuE Studio to Applications (notarized). For the phone, open app/YuERemote/YuERemote.xcodeproj in Xcode, set your team, and run on the device (xcodegen generate regenerates the project from project.yml).

YuE Studio 0.3.1: write lyrics with the on-device model

Write lyrics. A button beside the Lyrics label (shown when the Mac's on-device language model is available: macOS 26 with Apple Intelligence) asks what the song should be about, then writes verses, chorus, bridge and outro for that theme, the style and the title, in the generator's format, in a few seconds. A spinner covers the lyrics box while it works. It refuses with an alert while songs are queued or generating, since the model would take power from the GPU and Neural Engine.

Requires a Mac with Apple Silicon and macOS 14 or later. Download the DMG, drag the app to Applications, launch it and press Install. Signed and notarized.

YuE Studio 0.3: quality menu, titles, playback, transcription

Third release of YuE Studio for Apple Silicon.

Quality menu. Draft (GPU), Draft (GPU + Neural Engine), Full (GPU) and Full (GPU + Neural Engine). Draft is 8 solver steps, Full is 32; the engine part decides whether the scheduler may hand a song to the Neural Engine. Full (GPU) never maps the Neural Engine's 2.8 GB of weights, the low-memory choice for 16 GB Macs.

Titles. A run can be given a title, or left empty to have one chosen from the lyrics by the Mac's on-device language model (macOS 26 with Apple Intelligence; older systems use the first lyric line). The title names the run's folder and heads its rows.

Playback. A transport bar under the song list: play/pause (space), back to start, 10-second skips, scrubbing, elapsed and total time.

Transcribe a recording. A new button runs SheetSage2 (installed on demand, about 2 GB) to transcribe a source recording into a melody score for cover generation.

Also. A resizable style editor; two jobs submitted in the same second no longer share a run folder.

Requires a Mac with Apple Silicon and macOS 14 or later. Download the DMG, drag the app to Applications, launch it and press Install. Signed and notarized.

YuE Studio 0.2.1: Neural Engine stability with long songs

Fixes a failure with two long songs queued together.

Neural Engine. While one long song solved on the Neural Engine, the next song's program was compiled in the background and briefly loaded beside it, which overflowed the engine's per-process address window and failed the running song ("Program Inference error", status 0x2). Background compiles no longer load their program; a program is loaded only when its song takes the engine, after the previous one is unloaded. A pass that still fails with an inference error is retried once after reloading the program. Verified with two four-minute songs in different buckets.

Memory. The 8-bit token weights are quantized before each batch and dropped after it, and the MLX synthesis weights are dropped when no song needs the GPU for synthesis, keeping the worker's peak well below the machine's memory.

App. Stage progress is kept across status changes within a stage (the move to the Neural Engine no longer resets the bar). A stalled song (tokens saved, never synthesized) can be synthesized from its row, and doing so now actually starts: the worker rejected a song addressed by an audio file that didn't exist yet, leaving the row stuck on "Queued". A submit that fails for any reason now marks the song failed with the error instead of leaving it queued forever.

Also fixes importing the MLX synthesis module on machines without MLX installed.

Requires a Mac with Apple Silicon and macOS 14 or later. Download the DMG, drag the app to Applications, launch it and press Install. Signed and notarized.

YuE Studio 0.2: scheduler, 8-bit tokens, instrumental mode

Second release of YuE Studio for Apple Silicon.

Scheduling. Songs are processes and the GPU and Neural Engine are resources: one scheduler assigns them by arrival priority. Full-quality songs go to the Neural Engine; a song waiting for it starts on the GPU and moves across mid-solve when granted. Songs from different jobs tokenize together in one batch (up to four), each leaving the batch the moment its own tokens end. Rendering interleaves with tokenizing.

8-bit token decoding in MLX. Planning and tokenizing run in MLX with weights quantized to 8 bits: 18.5 ms per step against 33 to 37 on an M4, a 1.8x speedup at every batch size, same top choice and 0.9996 logit correlation with the PyTorch loop. Enabled on Macs with 24 GB or more.

Instrumental (no vocals). A switch that adds no-vocal tags, keeps only the section markers of the lyrics, and silences the vocal voice in the planned score before tokenizing.

App. Each song shows a Planning / Tokenizing / Synthing / Rendering track with live throughput; sections group songs by stage; the synthesis engine picker is gone (the scheduler chooses); a draggable splitter between songs and log.

Memory. Working copies (8-bit token weights, MLX synthesis weights) are dropped as soon as nothing needs them, keeping the worker's peak well below the machine's memory so the Neural Engine is never squeezed.

Requires a Mac with Apple Silicon and macOS 14 or later. Download the DMG, drag the app to Applications, launch it and press Install (the runtime and the 7 GB model download once; existing installs keep the model). Signed and notarized.

YuE Studio 0.1 for Apple Silicon

First release of the Apple Silicon fork.

YuE Studio is a native macOS app for YuE2 music generation. Download the DMG, drag the app to Applications, launch it and press Install: it sets up a private Python runtime and downloads the model weights (about 7 GB) once.

Requires a Mac with Apple Silicon and macOS 14 or later. 32 GB of memory is comfortable; 16 GB works with the lean loader.

Highlights: queued songs with per-stage progress and live throughput; tokens on the GPU, synthesis on the Neural Engine (or MLX for drafts), rendering on the GPU, all overlapping; a fast draft mode with full-quality re-render; an instrumental switch. Signed and notarized.