LM Studio's Locally App Lets iPhones Run 70B Models From Home

LM Studio just shipped its first official mobile app. Locally, the iOS chat app the company acquired earlier this year, is now the mainline LM Studio client for iPhone and iPad, and it ships with a feature called LM Link that lets your phone reach back into the desktop machine running your real models.

The pitch is simple. You probably have a Mac or workstation at home that can comfortably run a 30B parameter model. You also have a phone. Locally with LM Link stitches them together so the phone behaves like a thin client for the big box on your desk.

What LM Link actually does

LM Link is a feature in LM Studio that lets you connect together devices on which you have LM Studio (or llmster) installed. It is end-to-end encrypted, and built on top of custom Tailscale mesh VPNs. All devices in the LM Link network communicate with each other using a mesh VPN connection powered by Tailscale, using end-to-end encrypted connections without opening any ports to the internet. LM Link also runs entirely in userspace and does not change any global settings on your device.

The privacy story is the main reason this works at all for a local-AI crowd. LM Studio says chats remain on the local device and nothing gets uploaded to LM Studio's backend servers apart from your device list, which is used for discovery. Your conversations and the model weights themselves never round-trip through a third party.

On the iPhone side, the flow is: sign into your LM Studio account on both devices, activate LM Link, and the phone shows up in your mesh. LM Link works with any model users have installed on their Macs (including the built-in Apple Intelligence foundation model), with performance depending on the Mac's hardware, as with any local model.

It is not just a chat app

The interesting part for anyone building on top of LM Studio is that remote models behave like local ones everywhere, not only inside the Locally UI. Any model in your LM Link network can be used as if it is local. Any tool that already connects to your local LM Studio server will work with remote models too, just by pointing to localhost:1234 as usual. This means you can use LM Link models in tools like Codex, Claude Code, OpenCode, and any other tool pointing to LM Studio's local API.

In practice that means a laptop on hotel Wi-Fi can keep using the same OpenAI-compatible endpoint it always has, while the actual inference happens on a 128GB workstation back home. No code changes, no ngrok, no exposed ports.

Price and availability

Locally is live on the App Store now. LM Link is still gated. LM Link is in Preview, with access rolling out in batches. During the Preview period LM Link will be free. After that, in addition to a free plan, the company will release paid plans, with details still to come.

To get going you need LM Studio on the desktop, LM Link enabled on that account, and the Locally app on the phone. Pairing happens through the in-app flow.

Where it shines, where it stumbles

The strengths are predictable but real:

• Run 30B and 70B class models from a phone without renting GPU time
• Same OpenAI-compatible API surface as desktop LM Studio
• Mesh networking means it works on cellular, hotel Wi-Fi, or a coffee shop, not just your home LAN
• Chats persist on the device, not in a vendor cloud

The rough edges are the usual mobile-to-desktop tunnel problems. An early hands-on noted that the connection can get dropped too quickly after the iPhone app spends a few moments in the background, for instance when you leave the chat to find a document or do a quick web search, then hop back in. Throughput and latency also still depend on your home upload speed and the responsiveness of the host machine.

The bigger picture

This release closes the loop on a strategy LM Studio started earlier this year. The company announced its acquisition of the Locally app earlier in the year, and has been integrating Locally with LM Studio with the goal of adding a mobile surface for using AI models on the go. The acquisition gave them a native iOS codebase and Apple Silicon expertise. LM Link gave them the transport. Locally is the surface that ties the two together.

For anyone who has been running local models seriously, the awkward gap has always been the phone. You either gave up and paid for a hosted API the moment you stepped away from your desk, or you cobbled together Tailscale, a reverse proxy, and a third-party chat client like Chatbox or Enchanted. A first-party app from the same vendor that runs your desktop stack, with a tunnel designed for this exact job, is the kind of plumbing that should have existed already.