Limitations and unverified areas (version 1.0.0)
Never run on a physical iPhone or iPad
No device was connected during development. On iOS everything below was only compiled, or run in the Simulator where MLX cannot work: on-device MLX inference, memory behaviour under jetsam, thermal behaviour, the increased-memory-limit entitlement, background downloads, Face ID lock, dictation, the photo picker, haptics. The first TestFlight install is the first real test.
Not implemented
- GGUF. Local Chat proposes MLX models only. The llama.cpp backend remains on macOS for GGUF files imported by hand; it is not linked on iOS.
- Share extension. Files reach the app through "Open in Local Chat", the file importer, drag and drop and paste. There is no Share-sheet extension target.
- Speech transcription model. Voice input uses the system speech recognizer (on device when the OS supports the language, otherwise Apple's servers — the composer says which). No Whisper model is bundled.
- Structured output. A JSON schema sent to the local API is accepted and not enforced.
- Vision with GGUF and embeddings with GGUF.
- TLS on the local API server. It speaks plain HTTP; expose it on the LAN only on a network you trust (authentication is then mandatory).
- CloudKit sync. Sync goes through a folder you choose (for example iCloud Drive) or a Mac relay, encrypted end to end.
- Automatic updates of the macOS app.
Implemented but never exercised by hand or by a test
- The screens written last: Model Arena, Knowledge Studio, Agent Studio, Settings (server, nodes, sync, storage). They compile on both platforms and the Models, Observatory and Chat screens were opened and checked; these others were not clicked through.
- The local API server inside the app (the server module itself is tested over real sockets, including against the remote-node client).
- Sync between two real devices (the engine is tested with two stores through a folder and through the relay handler).
- Apple Foundation Models generation on a Mac.
- Vision models, MLX embedding models, tool calling with GGUF models.
- Gated Hugging Face repositories with a real token; background download sessions.
- The shell tool inside the sandboxed Release app (the sandbox restricts what a command can reach).
- VoiceOver, Dynamic Type, right-to-left layouts.
Known behaviour worth knowing
- Stopping a generation discards the KV cache of that conversation; the next turn re-evaluates the whole prompt.
- Titles are generated only by on-device models; with a remote model the title is the first line of the message.
- The live rate in the Observatory counts stream events until the backend's final token counts arrive.
- Parameter counts are "declared" when the Hub publishes them, otherwise "inferred" from the repository name, and labelled as such.
- No claim of superiority over other local AI apps is made: no comparative benchmark was run.