LM Studio Ships Locally + LM Link: Local LLMs Go Mobile

Running a local LLM on your iPhone just got a first-party answer: on June 4, 2026 LM Studio shipped version 0.4.16 with Locally, a native iPhone and iPad app, paired with LM Link — a remote-access layer that connects the phone to the models already installed on your Mac. The twist is that nothing routes through a cloud relay. The connection is end-to-end encrypted, chat history stays on each device, and the only data that reaches LM Studio's servers is the device discovery list used to find your machine.

That design choice is the whole story. Most mobile-AI coverage treats the world as a binary — either you send prompts to a cloud model (ChatGPT, Claude) or you run a tiny model natively on the phone's chip. LM Link is a third path that almost nobody articulates: your Mac becomes a private inference server you actually own, and the phone is just an encrypted terminal into it.

This guide separates the two products LM Studio shipped, walks through the Tailscale-based privacy architecture, explains the hardware reason the Mac has to be the compute engine, and lays out a clean three-way comparison so you can decide whether this belongs in your stack. Where figures come from LM Studio's own materials, we label them as vendor-stated and hedge accordingly.

LM Studio released version 0.4.16 on June 4, 2026, introducing the Locally mobile app for iPhone and iPad alongside LM Link remote-access technology. It is the visible payoff of an acquisition: on April 8, 2026, LM Studio acquired the Locally AI app and brought in its creator, Adrien Grondin, to lead native mobile development. The company framed the move as "doubling down on our mission of making AI accessible and useful to you, across your devices, wherever you go."

The release lands on top of a busy spring of incremental work. Version 0.4.13 (May 13) updated the MLX engine to v1.8.1; 0.4.14 (May 22) shipped MTP Speculative Decoding as stable; and 0.4.15 (May 29) added CUDA tensor parallelism for multi-GPU loading and fixed a prompt-cache bug that had been dropping the cache on every message when used with Claude Code. Locally and LM Link are the headline of 0.4.16, but they sit on a steadily maturing inference stack.

The easiest way to get confused here is to treat "Locally" and "LM Link" as one product with two names. They are distinct. Locally is the iPhone and iPad app you install from the App Store. LM Link is the remote-access protocol that lets that app — or other tooling — reach an LM Studio instance running somewhere else. Locally is the most visible consumer of LM Link, but LM Link is the reusable plumbing.

That distinction matters because LM Link is not limited to the phone. LM Studio ships a headless build called llmster for server environments without a GUI, and LM Link can connect to llmsterinstances too. In practice the "Mac" on the other end of the link can be a headless Linux server or a cloud VM running LM Studio's inference stack — not only a desktop with a monitor attached.

LM Link is also not the first attempt at this problem — it is the first-party answer to it. Community apps such as "Off Grid" and "LM Mini" had already solved the iPhone-to-local-LM Studio connection for early adopters. What LM Studio adds is an official, team-maintained, end-to-end encrypted implementation with no third-party relay in the path. If you want to understand where this fits in the wider movement toward an on-device private AI stack, LM Link is the consumer-facing edge of it.

The privacy claim is the part worth getting precise about, because "private" is overused in AI marketing. Here is what LM Studio states: the Locally app connects an iPhone to a desktop LM Studio instance over an end-to-end encrypted connection using custom Tailscale mesh VPNs. Devices are never exposed to the public internet, and Tailscale itself cannot see the data. The only thing that reaches LM Studio's servers is the device discovery list used to establish the peer-to-peer connection.

Chat history, by design, stays on each local device. That single fact is what separates LM Link from cloud-relay apps: there is no server in the middle accumulating your conversations. The discovery list exists only so the two devices can find each other; once they do, traffic flows directly between them.

The cryptography underneath is not exotic, which is reassuring rather than disappointing. Tailscale's WireGuard-based mesh uses ChaCha20-Poly1305 for encryption and Curve25519 for key exchange — industry-standard modern primitives. The coordination server only exchanges small encryption keys; the actual data traffic flows peer-to-peer. LM Link layers its own self-contained Tailscale usage on top via tsnet, so it can deliver this model without asking you to manage a Tailscale account yourself.

LM Link works with any model installed on your Mac, and performance depends on the Mac's hardware — because inference runs on the desktop, not the iPhone's chip. The phone is the input/output terminal; the Mac is the compute engine. That is not an arbitrary design decision. It is the only design that makes large models usable on a phone today, and the reason is memory bandwidth.

High-end iPhones built on the A17 or A18 Pro carry roughly 8 GB of unified memory and somewhere in the range of 50–90 GB/s of memory bandwidth. A Mac Studio with an M4 Ultra reaches up to 800 GB/s. Since LLM token generation is bandwidth-bound, that gap is the difference between a model that crawls and one that feels interactive. The phone simply cannot feed a 7B-or-larger model fast enough to be pleasant to use.

That bandwidth ceiling is why native on-device inference, without a Mac in the loop, is limited to small models. The practical floor for usable iPhone-native inference is models under 3B parameters with 4-bit quantization. An independent 2026 analysis reports a 125M parameter model running at roughly 50 tokens per second on current iPhones, with a 1B model practical for short-context tasks and the Apple A19 Pro Neural Engine delivering around 35 TOPS. Useful for a quick autocomplete or a small summarizer — not for the 7B-to-30B-class models people actually want to chat with.

This is the analytical core of why LM Link is interesting rather than redundant. The "use your Mac because the iPhone can't run big models" line shows up everywhere, but the bandwidth and RAM numbers explain why it is true, and why a relay-less bridge to your own hardware is the only way to get desktop-class local models onto a phone right now.

Most coverage frames mobile AI as cloud versus on-device. There is a third option, and the cleanest way to see the trade-offs is to put all three side by side. The table below is our own framing, with cells sourced from LM Studio's materials, the cloud providers' public pricing, and independent on-device analysis. Speed figures are deliberately rough and hardware-dependent.

Read down the privacy column and the value of LM Link becomes obvious: it is the only row that gives you device-local chat history and access to large models. Cloud trades privacy for capability; native trades capability for total offline privacy; LM Link keeps the conversation private while letting your Mac do the heavy lifting. The one thing it cannot do is run with no machine at all — if your Mac is asleep or off the network, there is nothing to connect to.

On cost, the math is straightforward. Cloud assistants run about $20 a month, or $240–480 a year. The LM Link and Locally software is free during preview, so the only spend is the Mac hardware you likely already own — a one-time sunk cost rather than a recurring subscription. For teams already weighing a self-hosted local LLM workflow, LM Link is a low-friction way to put that infrastructure in everyone's pocket.

The detail that makes LM Link more than a consumer toy is the API surface. LM Studio's local API is OpenAI-compatible, served on localhost:1234 by default. Any app or SDK that supports a custom OpenAI base URL can point at it — and that same API surface is what LM Link exposes remotely. So tools like Codex CLI, Claude Code, and OpenCode that already target that endpoint work over LM Link without reconfiguration.

That compatibility is the unlock. You are not adopting a proprietary protocol; you are taking the local endpoint you already use on your desktop and reaching it from elsewhere over an encrypted tunnel. For a developer, "my laptop's model, available from my phone or a second machine, on the API I already wrote against" is a much smaller change than it sounds.

One runtime nuance is worth a careful note. On Apple Silicon, vendor and community reports suggest MLX-format models run roughly 30–50% faster than llama.cpp on Metal, attributed to unified memory removing CPU-to-GPU transfer overhead. Treat that as an approximate, community-reported range rather than a controlled benchmark — no single published test pins down that exact figure, and your real speed depends entirely on the model, quantization, and the specific Mac doing the work.

A few things deserve to be stated plainly, because they are easy to gloss over in launch-day enthusiasm. First, LM Studio is not open source. It is free for home and work use, and it runs on macOS, Windows, and Linux — but its license is more restrictive than free/libre/open-source software, a point the developer community has raised. If open-source licensing is a hard requirement for your organization, treat LM Studio as a free tool, not an open one, and evaluate accordingly.

Second, LM Studio is not a model provider. It does not bundle or license models; it facilitates downloading and running third-party open-weight models — Llama 3.x, Qwen3.x, DeepSeek R1, Mistral, Gemma 4, Phi-3.5 and hundreds of community fine-tunes via Hugging Face integration. The quality you get is the quality of whichever open-weight model you load, not something LM Studio itself produces.

Third, on availability: LM Link is in a request-gated Preview, free for now, with free and paid tiers planned at general availability. The Locally app is available free during this Preview period, but permanent App Store pricing has not been confirmed in primary sources — so plan around "free at launch during preview," not a guaranteed-free future. And the launch is iPhone and iPad only; Android support has not been announced.

The rollout is deliberately incremental. The June 2026 Locally and LM Link launch was batch-limited, with preview access granted via a request form rather than an instant open release — consistent with how LM Studio has staged previous major features. The practical path is three steps: run LM Studio on the Mac, request LM Link Preview access, then install the Locally app and pair the devices.

For most individuals, the deciding question is simple: do you already own a Mac strong enough to run the model class you want? If yes, LM Link converts that hardware into a private, always-available inference server you reach from your phone — at no recurring software cost during preview. If you are weighing it against staying on a cloud assistant, it sits naturally alongside the broader decision of switching away from cloud chat apps. For teams, the same evaluation belongs inside a deliberate stack choice — exactly the kind of comparison our AI digital transformation engagements are built to run.