Generation, training, and stem separation run on your Mac. No server, no account, built on ACE-Step 1.5 and Apple's MLX.
Silo Systems builds infrastructure for people who would rather not negotiate with their own software. Ohm Studio is the same discipline applied to making music: a studio that runs entirely on the machine you already own.
Every generation, every training step, every analysis runs on your Mac via Apple's MLX framework. No server to talk to. No account to sign into. Models download once from public open-weights repositories, then operate locally for the lifetime of the app.
No analytics SDK, no telemetry, no crash reporter beyond Apple's opt-in macOS diagnostics. The PrivacyInfo manifest declares no tracking domains. Your prompts, lyrics, datasets, and generated audio stay in your Application Support folder. Exactly where you put them.
ACE-Step is an MIT-adjacent open-weights model. The audio you generate carries no per-stream licensing, no creator credits, no platform claim. Render it, sell it, sample it, throw it away. Ohm Studio doesn't appear in your release notes unless you put it there.
Type a prompt. Drop in reference audio. Add style tags. Pick a duration. The generation runs in MLX on your Apple Silicon GPU and writes a file to your Application Support folder. There is no submit endpoint. There is no job ID. There is no rate limit. The first generation downloads the model weights from the disclosed public repository after you approve the download; every generation after that is offline.
Drop a folder of audio into the Dataset tab. Choose a trainer: the fast compact backend for sketches, or the ACE-Step DiT LoRA/LoKR pipeline for serious work. Watch the loss curve. Cancel, pause, resume, delete; the job drawer is real. Trained adapters export as .ohmlora files. Bring them to another Mac without re-running the job.
Try the offline-first LoRA training simulator below. Set the hyperparameters (rank, alpha, learning rate), choose a dataset, and start the job to watch the loss curve converge.
Train a custom adapter (Low-Rank Adaptation) entirely on your device. Fine-tuning projects freeze the base model weight tensors, injecting compact rank-decomposed parameter blocks to map your private sound library onto ACE-Step 1.5 style vectors.
Ultimate Vocal Remover models run through the maintained audio-separator package. Vocals, drums, bass, instrumental: each written as a local file you can drag into any DAW. No upload step. No re-encoding. No watermark. When the optional UVR runtime isn't bundled, Ohm Studio falls back to local spectral stems with a typed message. Never a silent fake-track.
Local ACE-Step 1.5 inference. Lyrics, prompts, style tags, reference audio. No round-trip latency, no rate limits, no subscription throttle.
Train compact adapters on your own dataset. Selectable trainers from a fast numpy compact backend up to the full ACE-Step DiT LoRA/LoKR fine-tuning pipeline.
.ohmlora packaging. Bring trained adapters across machines, share them with collaborators, or version them in your project archive.
Ultimate Vocal Remover models run through the audio-separator package, written to local files you can drop into any DAW. Falls back to spectral stems when the optional UVR runtime is unavailable.
Real librosa BPM, key, chord, mood, genre, and structure analysis. Packaged-safe numba caching. Output that matches what your ear already knows.
Local mlx-lm runtime. Fails clearly when the local model is unavailable rather than silently falling back to the cloud. There is no cloud to fall back to.
Trim, reverse, time-stretch, fade, normalize, and export inside the app. Generated audio is editable before final export.
Record or import a short clip; the model takes it as a style/mood prompt. Microphone access is requested only when you use the feature.
No subscription. No usage tier. No credit purchase. One purchase, one app, with updates delivered through the App Store. Supported generation is available on Mac and iPad; training and separation remain full Mac studio workflows running on hardware you already own.
View on the App StoreNo. After you approve and complete supported model setup, Ohm Studio runs offline. Generation stays on supported Mac and iPad hardware; training, stem separation, and the full analysis studio stay on your Apple Silicon Mac. There is no hosted inference server or account.
Ohm Studio uses ACE-Step 1.5, an open-weights music generation model, running through Apple's MLX framework for native Apple Silicon acceleration. You can also train custom LoRA and LoKR adapters on your own music datasets.
Yes. ACE-Step is an MIT-adjacent open-weights model. The audio you generate carries no per-stream licensing, no creator credits, and no platform claim. You can use it commercially, sell it, sample it, or distribute it however you like.
The Mac studio requires macOS 14 or later on Apple Silicon. The touch-first iPad generation workflow requires iPadOS 17 or later and supported high-memory Apple Silicon hardware; model setup shows the exact device-fit result before download.
Suno and Udio are cloud-based services that process your prompts and audio on their servers. Ohm Studio runs 100% locally on your Mac. No cloud, no account, no data collection. You own the model, the adapters, and every piece of audio you generate. There are no usage limits, no subscriptions, and no rate limits.
Yes. Ohm Studio includes LoRA and LoKR fine-tuning pipelines. Drop a folder of audio into the Dataset tab, configure your hyperparameters, and train a custom adapter entirely on-device. Trained adapters export as .ohmlora files you can share or move between Macs.
Zero. No analytics SDK, no telemetry, no crash reporter beyond Apple's opt-in macOS diagnostics. The PrivacyInfo manifest declares no tracking domains. Your prompts, lyrics, datasets, and generated audio stay in your local Application Support folder.
$19.99 as a one-time App Store purchase for Mac and iPad. No subscription, usage tiers, or credit purchases.
Ohm Studio is a one-time $19.99 App Store purchase. No subscription, account, or cloud. Generate on Mac and iPad; use the full training and stem studio on your Apple Silicon Mac.
Vocal and stem separation, offline.
The practice room that never listens back. Slow any song, see its chords, loop the hard bars — 100% on your Mac.
Real-time vocal pitch correction and AI harmony.
An 8×8 AI groovebox for Mac and iPad.
