Spotlight: Local LLM — chat with AI on your iPhone, no cloud involved
Most AI chat apps are just a skin over an API call to someone else's server. Local LLM is the opposite bet: it downloads a model from an in-app catalog and runs it entirely on your iPhone. No account, no cloud chat history, no round trip to a server that could log your prompts.
What you actually get, based on the app's own description: a catalog of curated MLX models you download and run locally, text or voice chat, the ability to attach photos and files, and vision models when your device has the memory to spare. There's also a Live mode that adds hands-free voice conversation, using on-device speech recognition to listen and read-aloud replies to respond, so the whole loop stays on the phone rather than shuttling audio to a server.
The smart thing here isn't the model itself (MLX is Apple's own framework, and plenty of people are experimenting with it). It's the positioning: instead of promising to be smarter than ChatGPT, Local LLM is explicit that it's chasing a specific, narrower feeling, something closer to Gemini's on-device experience, but achieved through local inference rather than a cloud-first product with an offline mode bolted on. That's an honest way to frame a beta. It doesn't have to win on raw capability against GPT-4-class models it can't run on a phone. It has to win on the fact that nothing you type ever leaves the device.
That tradeoff is also the app's built-in limitation, and to its credit, the beta notes are upfront about it: large models need real storage and RAM, first load can be slow, and not every catalog model fits every device. Running a language model on a phone chip is still running a language model on a phone chip. If you're used to instant, GPT-4-tier responses, an on-device model on an iPhone, even a recent one, is not going to feel as sharp or as fast. That's the physics of it, not a flaw specific to this app.
Who should try it: people who care about privacy as a first-order feature, not an afterthought, iPhone owners who want to poke around with local models without setting up anything on a laptop, and anyone curious what current on-device LLMs are actually capable of today rather than in a demo video. If you've been mildly annoyed that every AI chat app wants your account and phones home your conversation history, this is aimed squarely at you. Indie hackers building offline-first or privacy-first products might also want to see how someone else solved the model-download-and-run problem on iOS, since that's genuinely fiddly to get right.
Who should skip it for now: anyone who needs consistently fast, high-quality answers for real work. This is a beta, distributed through TestFlight rather than the App Store, which means it's pre-release software with the caveats that implies, including the fact that Apple caps each TestFlight build at 90 days before you'd need a new one. If you have an older iPhone with limited RAM, several of the catalog models reportedly won't fit anyway, so check device requirements before you get excited. And if what you actually want is a bigger, smarter model for complex reasoning or coding help, on-device is the wrong architecture for that job right now, full stop.
Where I'd guess this goes next: the obvious pressure point is model selection and download experience, since that's where most local-LLM apps live or die for normal users who don't know what a quantized model is. If Local LLM can make picking the right model for your specific device close to automatic, and keep Live mode responsive enough to feel conversational rather than laggy, it has a real niche: the privacy-conscious chat app that doesn't ask you to trust anyone with your data. Getting out of TestFlight and onto the App Store, with a clearer answer on pricing, seems like the next obvious step, and it's the one that will tell us whether this is a hobby project or a real product.
Try Local LLM: testflight.apple.com
See the launch: Local LLM on welaunch.sh