Generation runs on your Mac. So does training. So does separation. There is no server. Not because we hide it; because there isn't one. Ohm Studio is a local-first AI music studio built around ACE-Step 1.5 and Apple's MLX framework.
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 ModelScope (with Hugging Face as a fallback); 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.
Interact with the live, offline-first LoRA adapter training simulator below. Tweak the hyperparameters (Rank, Alpha, and Learning Rate), choose a dataset, and start the job to observe gradient weight updates and loss convergence in real-time.
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, every future update included via the Mac App Store. Generation, training, separation: all unlimited because they all run on hardware you already own.
Coming to the Mac App Store ⌘Ohm Studio is in final pre-release for the Mac App Store. Join the waitlist and we'll send a single email when it goes live. No marketing list, no newsletter, no follow-up campaigns.
