Nativ Runs AI Models Locally on Apple Silicon
GitHub Pages ● Covered by 3 sources
Nativ runs open AI models straight on your Mac, no cloud or account needed. Unlike rival apps, its entire codebase is open source, not just the engine underneath.
Based on reporting by GitHub Pages — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nativ is a new desktop app built to run open AI models directly on Apple Silicon Macs, cutting cloud, subscriptions, and accounts out of the equation entirely. It ships with a curated library pulling standout open models from Google, Cohere, and Liquid AI, and it recommends which model actually fits your hardware instead of leaving you to guess.
Under the hood, Nativ leans on MLX-VLM and is tuned specifically for M-series unified memory and Metal, skipping the wrapper and translation layers that slow down other local tools. The chat interface mirrors what you'd expect from something like Claude, with streaming responses, markdown rendering, code highlighting, and image input. And it goes well past plain text chat: the app also handles vision, video summarization, code autocomplete, and speech transcription and generation, all running on-device.
For anyone who wants to see what's actually happening under the hood, Nativ surfaces live telemetry right in the chat window — tokens per second, memory pressure, thermal state, and time-to-first-token. It also connects to coding agents people already use, including Pi, Codex, Claude Code, Hermes, and OpenCode, acting as a single local model server that every one of those tools can plug into.
What sets Nativ apart, according to its own manifesto, is that the entire desktop app is open source and MIT licensed, not just the model engine underneath. The team is blunt about why: other "local AI" apps, they argue, are proprietary shells built on open-source engines they don't own, with closed UIs and paywalls hiding what's really running. Nativ says every line, every model loader, and every telemetry chart can be read, forked, or pulled against tonight — no VC roadmap, no enterprise tier, and no plan to quietly turn prompts into training data.
My take — AI-written commentary, not fact-checked reporting
Calling out other local AI apps for slapping a paywall on someone else's open-source engine takes nerve, but it's a fair shot, and it's genuinely rare to see a tool mean it when it says open source — UI included, not just the model weights. The manifesto is nice, but the real question is whether models from Google, Cohere, and Liquid AI run well enough on ordinary Macs to make people actually ditch cloud subscriptions. If the free-forever, no-data-harvesting promise holds up in practice, Nativ has earned the right to be a little smug about it.”
Read more about this at: GitHub Pages
