When we launched the first alpha of Stelix a few weeks ago, we shipped a macOS app that could run AI models locally on Apple Silicon. It worked. A small group of early testers used the alpha and gave us valuable feedback. But as we listened to what people were asking for and looked at where the technology is heading, we realized something: we could do this so much better.
Today we want to share what we have been working on. We are rebuilding Stelix from the ground up as a fully native Apple application, powered by Apple's MLX framework, and we are bringing it to the App Store for both iOS and macOS.
Why Go Native?
The first version of Stelix was built as a cross-platform desktop app. That approach let us move fast and validate the idea, but it came with compromises. The app did not feel entirely native on macOS. Window management, keyboard shortcuts, system integration; small things that add up. And when people started asking about an iPhone or iPad version, the answer was always going to require a fundamentally different approach.
Going native with Swift and SwiftUI means Stelix will feel like it belongs on your Mac, your iPhone, and your iPad. It means proper system integration, smooth animations, correct platform conventions, and the kind of polish that users expect from apps they use every day. It also means we can take full advantage of Apple's hardware and software stack in ways that were not possible before.
What Is MLX and Why Does It Matter?
MLX is Apple's machine learning framework, built specifically for Apple Silicon. If you have not heard of it, think of it as Apple's answer to PyTorch, but designed from scratch for the M-series chips and the Neural Engine.
What makes MLX special for Stelix is how it handles memory. On Apple Silicon, the CPU and GPU share a unified memory architecture. MLX is built to exploit this. Instead of copying data back and forth between CPU memory and GPU memory like traditional frameworks do, MLX lets the GPU access model weights directly from shared memory. This means larger models can fit on devices with less RAM, inference starts faster, and the whole pipeline is more efficient.
For a 7-billion parameter model running on an M2 MacBook Air, the difference is meaningful. Responses start faster. Token generation is smoother. And because there is less memory pressure, you can keep Stelix running alongside your other apps without everything grinding to a halt.
On iPhone and iPad, MLX works alongside the Neural Engine, Apple's dedicated machine learning accelerator. This opens the door to running smaller but still capable models on mobile devices, something that would have been impractical just a year ago.
Running AI on Your iPhone
Let's talk about what this actually means in practice. With MLX and the latest A-series and M-series chips, we can run language models with billions of parameters directly on your phone. Not a stripped-down version. Not a cloud fallback with a local cache. Actual on-device inference.
The models we are targeting for iOS are carefully selected for the best balance of quality and performance. They are quantized and optimized to run within the memory and thermal constraints of mobile hardware, but they are still remarkably capable. You will be able to have natural conversations, brainstorm ideas, draft text, and ask questions, all without an internet connection.
Imagine pulling out your phone on the subway and having a full AI assistant available. No cell signal required. No loading spinners while a server processes your request. Just instant, private AI that works wherever you are.
The App Store Changes Everything
Distributing Stelix through the App Store is a big shift for us, and we think it is the right one. Here is why.
Easier access. No more downloading a DMG file and dragging it to your Applications folder. No more macOS Gatekeeper warnings. You search for Stelix, tap install, and you are done. On iPhone, it is the only way people expect to get apps.
Automatic updates. When we ship improvements, you get them automatically. No need to check our website, download a new version, and replace the old one. The App Store handles it.
Trust and transparency. Apple reviews every app and every update before it goes live. That review process means you can trust that the app does what it says and nothing more. For a privacy-focused app like Stelix, that third-party verification matters.
One download, all devices. We are building Stelix as a universal app. That means one download works on your Mac, your iPhone, and your iPad. Your experience travels with you across Apple's ecosystem.
What Happens to the Alpha?
We have taken down the alpha download. Rather than maintaining an older version while building the new one, we decided to focus all of our energy on getting the App Store release out as soon as possible. We want new users to experience Stelix at its best from the very first launch, and the native App Store version will be a significant step up from where the alpha left off.
We are grateful to everyone who tested the alpha and shared their feedback. Those early conversations shaped many of the decisions we are making now.
The Technical Stack
For those who are curious about what is under the hood, here is a high-level look at the new architecture:
Swift and SwiftUI for the entire application layer. Native views, native navigation, native lifecycle management. The alpha was built with a cross-platform framework, which served us well for rapid prototyping, but that approach will always have limits when it comes to deep platform integration. Going fully native removes those limits.
MLX for all model inference. The alpha used a portable, cross-platform inference engine under the hood, which got the job done but was not optimized for Apple hardware. MLX is built specifically for Apple Silicon and takes full advantage of the unified memory architecture. We are using MLX's Swift bindings directly, which means the inference pipeline is written in the same language as the rest of the app. No bridging layers, no interprocess communication overhead.
Core Data for local conversation storage. Your chats stay on your device, encrypted at rest by the operating system. iCloud sync is something we are evaluating for the future, but only as an opt-in feature and only with end-to-end encryption.
Metal Performance Shaders as a fallback for specific operations where we need lower-level GPU control. Most of the heavy lifting is handled by MLX, but having direct Metal access gives us room to optimize hot paths.
What Comes Next
We are deep in development right now and making progress every day. We do not have an exact release date to share yet because we want to get this right, not just get it out. Rushing a native rebuild would defeat the purpose.
Here is what we can tell you about the roadmap. The macOS App Store version will come first, since that platform is closest to what we have already built and tested. The iOS version will follow shortly after, sharing the same core engine and data layer.
We will keep sharing updates here on the blog as we hit milestones. If you want to know the moment Stelix lands on the App Store, keep an eye on this space.
The Bigger Picture
We started Stelix because we believed people deserve an AI assistant that respects their privacy. That has not changed. What has changed is our understanding of how far we can push that idea.
With MLX and native Apple development, we are not just making a privacy-friendly alternative to cloud AI. We are building something that is genuinely better for a large number of use cases. Faster startup, zero latency, works offline, no account required to get started, and deep integration with the platform you already use.
The future of AI is not only in massive data centers. It is also in your pocket. And we are building Stelix to be right there with you.