The local-AI-on-Mac space has been fragmented for a while: Ollama for text, separate wrappers for multimodal, and llama.cpp doing its best on hardware that was never designed around it. Nativ is an attempt to consolidate that into a single, opinionated SwiftUI app — and its author has the right credentials to pull it off.
Prince Canuma is the maintainer of MLX-VLM, the library that powers multimodal inference on Apple Silicon in LM Studio and several other tools. His track record here is real. Nativ v0.0.1 shipped on July 20 and already has 300+ GitHub stars after a day on Hacker News; it’s written almost entirely in Swift and requires Apple Silicon with macOS 26.
The inference backend is MLX, Apple’s JAX-inspired framework that runs natively on the unified memory architecture and Metal GPU. On M-series chips, MLX consistently outperforms llama.cpp for multimodal workloads, and because both the CPU and GPU share the same memory pool there’s no copy overhead. Nativ exposes this through a model manager that pulls open weights from Hugging Face — currently Google’s Gemma 4 E2B (10 GB, vision+audio), Cohere’s North Mini Code (19 GB, tools), and Liquid AI’s LFM2.5-VL 1.6B (3 GB) — with the usual caveats that a 19 GB model download is nontrivial and your M2 Air will behave differently from an M4 Ultra.
What distinguishes Nativ from plain Ollama or the existing MLX demo scripts is the server layer. The app runs an inference server that speaks both the OpenAI Chat API and Anthropic’s Messages API, which means you can point Claude Code, Cursor, or any tool that talks to those endpoints at localhost and use a locally-hosted model with zero API key overhead. That’s not new as a concept — llama.cpp has had an OpenAI-compat server for years — but doing it with MLX multimodal models in a packaged macOS app is cleaner than running scripts in a terminal.
There’s a performance dashboard built in too: tokens per second, memory pressure, and thermal state shown live. That’s genuinely useful when you’re trying to figure out whether you’re bottlenecked by model size or thermal throttling, and whether a model is actually worth keeping around at its download size.
The roadmap mentions audio-only and image-generation-only models, which suggests the author intends to cover more of the MLX-Audio surface he also maintains. The pieces fit: MLX for computation, Swift for the native shell, and the existing Hugging Face model ecosystem for weights.
A few practical notes for anyone setting it up: the app requires macOS 26 (the current release), Apple Silicon is mandatory, and the bigger models will saturate unified memory on 8 GB and 16 GB machines. The OpenAI-compat server mode needs a quick toggle in settings. If you already have MLX-VLM installed separately, Nativ can find models in your existing Hugging Face cache rather than re-downloading them.
It’s a v0.0.1, so surface polish and edge-case handling will need work. But the architecture is sound, the author knows the problem space, and having a single native app that covers local chat, inference server, model management, and performance monitoring on Apple Silicon is something the ecosystem has been missing.