MLX vs FrankenWhisper

MLX-Serve is a private AI studio that runs entirely on your iPhone. Every conversation, voice recording, and generated image is processed on-device by open-weight models — nothing you type, say, or create ever leaves your phone. There is no account to create, no subscription, and no tracking of any kind. MUSIC Generate songs locally using your device ! You can also play it back in the background and use MLX-Serve as a player for your generated music on the go. CHAT Talk with capable open language models (Google Gemma 4) right on your device. Ask questions, brainstorm, draft messages, get things explained. Some models can even understand photos you attach. CALL Tap Call inside any conversation to talk to the AI hands-free, like a phone call. Your speech is transcribed on-device and the reply is read back aloud. Speaker and handset modes are both supported, and the transcript is saved to your chat. VOICE CLONE Record about eight seconds of your voice and have the AI speak text in your own voice, using an on-device text-to-speech model. Great for narration and fun. IMAGES Generate images from a text prompt with an on-device diffusion model (FLUX.2). Tap a result to zoom in, save it to your photo library, or share it. FULLY PRIVATE, FULLY OFFLINE - Runs 100% on-device using Apple MLX and the Metal GPU. - No account, no analytics, no ads, no tracking. - The network is used only to download the AI models you choose from Hugging Face. After a model is downloaded, that feature works offline. REQUIREMENTS Models are multi-gigabyte downloads and on-device AI is demanding, so MLX Chat requires a recent, high-performance iPhone. Download over Wi-Fi is recommended. You pick which models to download and can delete them anytime to reclaim space. Open models are provided under their respective licenses (Gemma, Apache 2.0). MLX Chat is an independent app and is not affiliated with Google or Apple.

Turn recordings and videos into useful text without uploading them. FrankenWhisper transcribes microphone recordings, audio files, and playable videos from Photos, Files, or the share sheet. The full-quality lane uses the large-v3-turbo model. Live dictation and the optional custom keyboard use a smaller model for low latency, with local noise reduction in the real-time path. Speaker identification can separate up to four voices. Add names when you know them or leave them blank and use color alone. Export the transcript as text, Markdown, HTML, SRT, or JSON with timestamps and speaker structure. Subtitle Studio creates a new H.264 video with customizable captions, font, size, position, and color. Word-level timing drives a karaoke highlight, and named speakers can receive distinct labels and color schemes. All transcription, speaker analysis, denoising, timing, and video rendering run on the device. Model downloads are the only required network activity. The keyboard's Full Access setting is used only for the local App Group handoff between the keyboard and FrankenWhisper; the keyboard extension has no network client.