Nativ setup on Mac: models, dictation and Pi

Prince Canuma walks through Nativ, an open source Mac app for downloading models and using them through a graphical interface. The tutorial covers permissions, Hugging Face authentication, model discovery and memory warnings. The host uses an M3 Mac with 128 GB of RAM; Canuma says the smaller LFM 2.5-2.6B model should also run on a 16 GB device. Gated models require approval through Hugging Face before download.

The audio workflow pairs Cohere Transcribe for speech recognition with LFM for meeting summaries. The speakers demonstrate a recording library and global voice dictation, presented as a replacement for Whisper Flow. Canuma says these workflows can operate offline after the models download. He qualifies the private processing claim: external tools that use the internet can send data beyond the device.

Nativ also exposes a local server for other applications. The walkthrough configures Pi to use a local LLM, then explores GitHub MCP access and scheduled tasks. Exa supplies web search through an API key. The demonstrations include failed web-read calls, repository permission problems and an unsupported model format. Flux image generation produces several results that miss the requested subject.

The later discussion covers MLX Audio Swift for iPhone and iPad apps. Video input in Nativ's chat, speculative decoding and SSD streaming are described as upcoming features, rather than capabilities demonstrated in the app.