The mlx-vlm Maintainer Just Came for LM Studio's Lunch
Nativ is an MIT-licensed macOS app from the developer whose inference code already powers rival local-AI frontends.
There's a specific kind of grudge match brewing in local AI, and Nativ is the clearest shot fired yet. The new macOS app, which hit the Hacker News front page this weekend on the day of its v0.0.1 release, comes from Prince Canuma — the developer behind mlx-vlm, mlx-audio, and a stack of other inference libraries for Apple's MLX framework. Nativ's landing page doesn't name names, but the subtext is loud: "The other 'local AI' apps you've heard of? They're proprietary shells built on top of open-source engines they don't own."
Here's what makes that jab land harder than the usual open-source posturing: when LM Studio open-sourced its MLX engine in 2024, it publicly credited mlx-vlm as one of the packages it was built on. The closed-source app that dominates Mac-based local inference runs, in part, on Canuma's code. Nativ is the engine maintainer deciding to ship the whole car.
The vertical integration play
The pattern is familiar from the rest of the infrastructure world: the maintainer of a widely-embedded library watches downstream products capture the value, then goes direct. Nativ is a native SwiftUI app (about 85% Swift by the repo's own accounting) that bundles an mlx-vlm server, pulls compatible models from Hugging Face, and wraps it all in chat, model management, and a menu-bar control surface. MIT license, no accounts, no telemetry phoning home, free.
The honest caveat to the "proprietary shells" framing is that it really only describes LM Studio. Ollama is MIT-licensed end to end, and smaller players like Jan are open too. What Nativ can uniquely claim is that the app author and the engine author are the same person — when a new model architecture drops, the person patching mlx-vlm for day-zero support is also the one shipping the app update. That's been Canuma's signature move for years (mlx-vlm and mlx-audio routinely land same-day support for new releases), and it's a genuinely hard advantage for a shell app to replicate.
What's actually in the box
Two things in the feature list matter more than the chat UI. First, Nativ exposes local APIs in both OpenAI-compatible formats and Anthropic Messages format. Most local runners stop at OpenAI compatibility; shipping an Anthropic-shaped endpoint means you can point Claude Code at a local model by overriding ANTHROPIC_BASE_URL, no proxy shim required. Nativ leans into this with documented integrations for coding agents — Claude Code, Codex, OpenCode, and others.
Second, the model curation is aggressively current. The launch catalog centers on three releases from the past few months: Google's Gemma 4 E2B (April 2026, native audio and vision, about a 10 GB download), Cohere's North Mini Code (June 2026, a 30B mixture-of-experts coder with 3B active parameters and a 256K context window, roughly 19 GB quantized), and Liquid AI's compact LFM2.5-VL vision model at around 3 GB. North Mini Code is the interesting one for the agent workflow: an Apache 2.0 coding model designed for tool use, running on your own machine, fronted by an Anthropic-compatible API. That's a fully local Claude Code loop — a thing people have been duct-taping together with LiteLLM proxies for a year.
The app also surfaces per-message tokens-per-second, memory pressure, and thermal state. That sounds like dashboard candy until you've watched an M-series laptop quietly throttle halfway through a long agent session and wondered why generation crawled.
Where the skepticism is earned
The HN thread was rougher than the star count suggests, and the critics have a point on two fronts.
"Frontier" is marketing. Gemma 4's 31B flagship ranks among the top open models on leaderboards, but the E2B variant Nativ ships is an efficiency play for on-device use, and North Mini Code's own launch benchmark (33.4 on the Artificial Analysis Coding Index) positions it as competitive among similarly sized models. These are excellent small models. Nobody's local M2 Air is running anything a reasonable person calls frontier, and the title invites a comparison the products can't win.
The category is also brutally crowded. Commenters rattled off LM Studio, Ollama, Jan, Open WebUI, and oMLX — the last of which one commenter called "more complete and fleshed out" than Nativ today. Fair. This is a v0.0.1 released this weekend, it requires macOS 26 or newer (a real adoption cut — plenty of developers are still a version or two behind), and there's no Intel or non-Mac story at all, by design.
Should you care?
If you want a local runner that works today with the broadest hardware support, LM Studio and Ollama remain the pragmatic defaults; nothing Nativ shipped this weekend changes that. Switching now buys you an auditable MIT stack and not much else you can't already get.
But two groups should watch this closely. If you're wiring coding agents to local models — especially Claude Code, given the native Anthropic endpoint — Nativ is attempting the cleanest version of that workflow anyone's shipped, and North Mini Code on a 36 GB-plus Apple Silicon machine is a credible local backend for it. And if you care about the supply-chain question — whether the local-AI apps winning mindshare should be closed shells over open engines — Nativ is the test case for the alternative. The engine layer of Mac-local AI has been open all along. Whether the maintainer of that layer can out-ship the polished proprietary frontends built on top of it is the actual story here, and it's a better one than "frontier models on your Mac."
Sources & further reading
- Nativ: Run frontier open models locally on your Mac — blaizzy.github.io
- Blaizzy/nativ — github.com
- Nativ: Run frontier open models locally on your Mac (discussion) — news.ycombinator.com
- LM Studio: our MLX engine is open source (MIT), built with mlx-lm, outlines, and mlx-vlm — x.com
- Gemma 4: Byte for byte, the most capable open models — blog.google
- North Mini Code: Agentic Coding Model for Developers — cohere.com
Rachel has been embedded in the developer tooling ecosystem for nearly eight years, covering everything from IDE wars and package-manager drama to the quiet rise of AI-assisted coding. She has a soft spot for open-source maintainers and an unhealthy number of terminal emulators installed on a single laptop.
Discussion 5
oh damn, open source inference ui built by the mlx-vlm guy. weekend project energy but probably actually solid
yeah but what's the actual hosting/compute story here? mlx-vlm is incredible for apple silicon but i'm still fuzzy on whether nativ handles quantization/model downloads itself or if you're managing that separately. like does it bake in a way to actually get models on your machine or is that still a manual thing
been running mlx stuff for months and nativ's got the model hub baked in pretty cleanly—downloads and quantizes on first run, which saved me a lot of manual onnx conversion work. the real win for me was watching my first inference go from 800ms cold to 45ms warm on m3 max. what still gets me is whether they're doing any p99 tail latency work since the first couple runs hit memory pressure pretty hard.
honestly the fact that they're not being super clear about the model management layer in the launch materials is kind of telling — and yeah, i'd need to dig into the repo to be sure, but just having best-in-class inference doesn't automatically mean they've solved the ux problem of "ok but how do i actually get llama-2-13b on my mac without feeling like i'm building something." mlx-vlm is technically brilliant but it's always felt more like a building block than a complete story to me.
solid inference underneath doesn't solve the actual pain point though - GPU memory management and quantization strategy get real complicated once you're juggling multiple models or doing any kind of batching. wonder if nativ handles fallback gracefully when you're trying to load something that barely fits, or if it just tanks like everything else does
