The MLX Crowd Finally Gets Its Own LM Studio
mlx-vlm maintainer Prince Canuma ships Nativ, an open-source Mac app built to feed local models to coding agents.
If you've run a vision model on a Mac in the last two years, there's a decent chance Prince Canuma's code did the actual work. His mlx-vlm library is the de facto multimodal inference layer for Apple's MLX framework — the thing wrappers wrap. This week he shipped a wrapper of his own: Nativ, an MIT-licensed macOS app that hit v0.0.1 and the Hacker News front page on the same day.
The obvious question — does the world need another local-LLM desktop app? — has an obvious answer: no. LM Studio, Ollama, Jan, Open WebUI, and a half-dozen smaller MLX frontends already exist. But Nativ is worth a closer look anyway, for reasons that have less to do with the chat window and more to do with who built it and which API it speaks.
The plumbing finally got a first-party faucet
The MLX ecosystem has had an odd shape for a while. Apple's framework delivers excellent inference on M-series unified memory, and libraries like mlx-lm and mlx-vlm are actively maintained. What's been missing is a polished, open GUI that treats MLX as the primary format rather than an afterthought. LM Studio runs MLX models well, but the app itself is proprietary and Electron-based. Ollama's world is GGUF and llama.cpp; native MLX support has been a perennial ask. So MLX users have mostly lived in the terminal, or in thin community frontends of varying freshness.
Nativ is the library maintainer closing that gap himself: a Swift/SwiftUI app (about 85% Swift by the repo's own accounting) that bundles an mlx-vlm server and puts a model library, a multimodal chat UI, and live telemetry — tokens per second, memory pressure, thermal state, time to first token — on top of it. That telemetry panel is a small thing that signals the intended audience. Casual chat users don't care about thermal throttling. People deciding whether a 30B quant fits their 36GB MacBook do.
The first-party angle matters for a practical reason: every third-party MLX frontend eventually lags the inference library it wraps, because model support lands in the library first. When the same person owns both layers, a new architecture landing in mlx-vlm and appearing in the app is one release cycle, not two projects' worth of issue-tracker latency. That's the strongest structural argument for Nativ over the field, and it's not something LM Studio can easily copy.
The tell is the Anthropic endpoint
Nativ exposes local API endpoints that are OpenAI-compatible and Anthropic-compatible. The OpenAI-compatible part is table stakes — everything from llama.cpp's server on up does it. The Anthropic Messages API is rarer, and it exists for exactly one workflow: pointing coding agents at localhost. Claude Code honors an ANTHROPIC_BASE_URL override; Codex-style CLIs and OpenCode speak the OpenAI shape. Nativ's landing page names those agents explicitly.
That's the actual product thesis here, and it's well timed. Mid-2026 is the first moment where a local coding-agent backend is a defensible idea rather than a party trick. North Mini Code, Cohere's June release and one of Nativ's three launch partner models, is a 30B mixture-of-experts with roughly 3B active parameters, Apache 2.0, built specifically for agentic coding and terminal work — and a ~19GB quant of it runs on a well-specced MacBook. The other partners point at the multimodal edge: Google's Gemma 4 E2B with vision and native audio input, and Liquid AI's LFM2.5-VL at 1.6B parameters and about 3GB. None of these existed nine months ago in anything like their current quality.
Whether you should run your coding agent against a 3B-active model is a different question. For refactors across a real codebase, the gap to hosted frontier models is still wide, and agents amplify model weakness because errors compound across turns. Where it does hold up today: offline work, privacy-constrained repos, high-volume mechanical tasks (test scaffolding, docstrings, commit summaries) where you'd rather burn watts than tokens, and as a free fallback when you've hit a rate limit mid-flow. That's a real niche, not a revolution.
What to be skeptical about
Start with the word "frontier" in Nativ's own pitch, which HN commenters rightly picked apart. Gemma 4 E2B and LFM2.5-VL are impressive edge models. Calling anything that fits in 20GB of unified memory "frontier" is marketing, and it slightly undercuts an otherwise refreshingly honest project — the manifesto's actual differentiators (MIT license, no accounts, no telemetry-as-in-tracking, community ownership) stand fine on their own.
Then there's maturity. This is v0.0.1 from a repository that was hours old when it hit HN. The app requires macOS 26, which rules out a large slice of otherwise-capable Apple silicon machines that haven't upgraded. And "native" deserves an asterisk: the UI is Swift, but inference runs through a bundled Python mlx-vlm server. That's a sensible architecture — you get the mature library instead of a rewrite — but it's closer to LM Studio's app-plus-engine split than the branding suggests. A future all-Swift inference path (Canuma also maintains an mlx-audio Swift port) would make the iPhone/iPad story interesting; today it's aspiration.
Finally, MLX-first means Hugging Face's mlx-community quants are your catalog, not the vastly larger GGUF universe. For popular models that's a non-issue; for long-tail fine-tunes it will bite.
Worth a slot in the dock?
If you're on an M-series Mac running macOS 26 and you've wanted LM Studio's convenience without the closed source — or you want to point Claude Code at a local model without hand-rolling a proxy — Nativ is the most credible new entrant in a while, precisely because it's the infrastructure maintainer moving up the stack. If you're happy with Ollama and GGUF, nothing here forces a switch, and a one-day-old repo hasn't yet earned your daily driver. Check back at v0.2. Given who's writing it, there's a better-than-usual chance it gets there.
Sources & further reading
- Nativ: Run frontier open models locally on your Mac — blaizzy.github.io
- Nativ: Run frontier open models locally on your Mac (discussion) — news.ycombinator.com
- Blaizzy/nativ — github.com
- North Mini Code: Agentic Coding Model for Developers — cohere.com
- Gemma 4: Byte for byte, the most capable open models — blog.google
- Introducing LFM2.5: The Next Generation of On-Device AI — liquid.ai
Priya covers AI frameworks, developer productivity tooling, and the startup ecosystem across South and Southeast Asia, bringing a researcher's rigour and a practitioner's empathy to every story. She is deeply sceptical of benchmarks and asks hard questions so her readers don't have to.
Discussion 6
spent three weeks debugging vision model inference on an m3 mac before mlx-vlm existed—would've saved me days if something like Nativ had been around then. the real win here isn't another ui, it's having the maintainer who actually understands the inference layer shipping something opinionated for their own users instead of fighting ten different wrapper assumptions.
cool that it's optimized for mlx, but anyone actually running agents through this in production needs to measure what happens when the model OOMs or thermal throttles mid-inference. mlx on gpu is fast until it isn't — would love to see what the p99 latency looks like under sustained load vs the cherry-picked benchmarks we always see.
totally agree — the cheerful demo always stops right before reality hits. what really worries me is the thermal throttle cascade on sustained multi-turn agentic workflows. you get one OOM and suddenly the whole inference pipeline becomes unpredictable. mlx's memory efficiency story is real but it doesn't account for how aggressive mac thermal management gets, especially if you're trying to keep the fan quiet. someone needs to actually stress-test this with back-to-back agent calls and publish the p99 numbers, not just single-inference latencies.
mlx-vlm maintainer shipping their own app actually makes sense—fewer impedance mismatches. curious if it handles multimodal workflows better than the generic wrappers
yeah, that's the real sell. when the person who owns the inference layer also builds the UI they get to stop translating between abstractions. though curious how much of this is just solving their own workflow versus something that generalizes
that's a good point about the impedance mismatches—do you know if nativ actually exposes any mlx-vlm specific tuning options, or does it just wrap the standard inference path? like i'm wondering if prince could've baked in optimizations that the generic wrappers miss, or if it's mostly just cleaner ux around the same underlying calls.
