The setup in one paragraph
Raj says his desktop runs LM Studio v0.4.28, with Google's official 4-bit quantization-aware-trained (QAT) build of Gemma 4 26B A4B. He reports that all 30 layers sit on the GPU with a 16K context window, and that LM Studio showed a "Full GPU Offload Possible" badge before he downloaded the model. His phone runs Locally, which signs in with the same LM Studio account and lists the PC's model. Those are his own configuration details, not independent benchmarks. He gives no speed figures, so this article doesn't either.
Why a phone can't do this alone
Google's Gemma 4 documentation explains the mismatch. The 26B A4B is a Mixture-of-Experts model that activates only 4 billion parameters per token, but Google says all 26 billion parameters still have to be loaded into memory. That makes its memory needs resemble a dense 26B model's, not a 4B model's. Google's approximate estimates are 57.7 GB at BF16, 28.8 GB at 8-bit and 14.4 GB at 4-bit.
Google adds two caveats that matter for anyone buying hardware:
- The figures cover static weights plus a 20% loading allowance. They don't include the extra VRAM for supporting software or the context window.
- The KV cache grows with prompt and response length, so larger context windows need more VRAM on top of the weights.
So 14.4 GB doesn't promise that a 16 GB card fits every configuration. Raj's 16K context on a 16 GB-class card is plausible, but it's one reported setup, not a guarantee.
On quality, Google describes QAT as simulating quantization during training so the model learns to compensate for lower precision. It says QAT checkpoints perform nearly identically to their high-precision baselines. That is Google's claim, not a verified result for every task. Google also lists the qat-q4_0-gguf format for llama.cpp and LM Studio use, and says it is available for 26B A4B.
How LM Link works
LM Studio's documentation describes LM Link as a way to securely connect your devices and use your local models remotely. Locally is the LM Studio mobile app for iPhone and iPad. LM Studio announced the iPhone and iPad version on June 4, 2026. App Store version history shows LM Link support arriving in Locally 1.57.0 on the same date.
Setup steps
- Install LM Studio on the host computer and sign in to your LM Studio account.
- Make sure LM Link is enabled in LM Studio.
- Install Locally on the iPhone or iPad.
- Sign in to Locally with the same account and enable LM Link there.
- The devices discover each other over the link, and you can use the models running in LM Studio from the phone.
A separate DEV Community walkthrough adds a practical tip: keep LM Studio open and the model loaded on the host. It also recommends testing with a small model first.
What LM Studio says about security
- No public exposure: the FAQ says LM Link requires an internet connection to establish the connection, and that linked devices communicate over Tailscale's end-to-end encrypted connections. None of them are exposed to the public internet.
- Limited access: linked devices can't do anything beyond LM Studio model and API access. The FAQ says LM Link does not expose the operating system, files or other services.
- Hub visibility: the LM Studio Hub only handles discovery. Chats and model listings travel inside the encrypted connection, and the Hub can't see chats.
- Tailscale coexistence: LM Link is described as separate from any Tailscale VPN you already use. It also can't reuse an existing Tailscale network, because LM Studio builds a dedicated one.
One nuance: "local" describes where inference happens, not an absence of network traffic. The documentation says LM Link needs internet access to connect, and discovery goes through an account. Vucense makes the same point: LM Link isn't fully local, because discovery relies on a device list on LM Studio's servers. Your prompts and documents stay on your own hardware, but your devices still talk across the internet.
The cellular test
Raj turned off his phone's Wi-Fi and still reached the model over mobile data, without opening any router ports. LM Studio's documentation is consistent with that. It doesn't require the phone and PC to share a network, only an internet connection to establish the link. The test is his own and hasn't been independently replicated.
His comparison to a bare llama.cpp server is fair as far as it goes. Remote access to llama-server generally means setting up a VPN or exposing the server. LM Link bundles a Tailscale-based tunnel to avoid that. If you already run your own VPN, that advantage is smaller.
Limits to know before you rely on it
The host PC has to be awake. Raj notes that if Windows puts the desktop to sleep, the phone loses the model. That follows from the design: Locally is a client, and the inference happens on the PC. If you want the model reachable all day, check your Windows power and sleep settings. Keep in mind that this means leaving a GPU machine on.
Backgrounding can break the connection. LM Studio's documentation says there are limits on how long the encrypted connection survives once Locally is in the background. It says it is working on faster reconnection and ways to keep the link alive. Later Locally releases mention improved reconnection and a faster LM Link. Even so, don't expect a long chat to survive being switched away from indefinitely.
It's iPhone and iPad only for now. Official documentation covers iPhone and iPad. Third-party coverage says Android hasn't been announced. The App Store listing also shows Mac compatibility, but LM Studio says LM Link connects LM Studio on a computer to Locally on a mobile device. It doesn't link two Locally apps together.
Cost is in flux. Raj says LM Link is free only during its Preview and that LM Studio plans paid tiers. Vucense reports the same: free during Preview, with free and paid plans planned at general availability. I found no primary LM Studio source confirming pricing details, so treat future pricing as unconfirmed. Today's free status is the only part both sources agree on.
Minor discrepancies. Raj puts the Locally download at 116 MB. The current US App Store listing shows 124.6 MB and requires iOS 18.1 or later. Sizes shift with app versions, so don't read much into it.
Beyond the iPhone
LM Studio's FAQ says any model on your LM Link network can be used as if it were local. Tools that already point at localhost:1234 can use remote models without changes, and the API and SDKs work the same way. LM Studio's LM Link page also lists CLI, REST API and coding-tool integrations. That makes it interesting for developers running a heavy model on a desktop and working from a lighter laptop.
Is it worth trying?
If you already have a GPU that fits a 14 GB-class model and an iPhone, this is a low-effort experiment. The account sign-in, the toggle in LM Studio and the Locally app are all there is to it. The reported benefits are privacy-minded: your documents and prompts go to your own PC and not a vendor's AI servers.
It isn't a drop-in replacement for everything. You need an always-on PC, a stable connection, patience with background reconnects, and tolerance for the open questions about future pricing. Compared with cloud assistants, you give up convenience in exchange for control of the hardware. The App Store listing's own reviews are positive, but they are user opinions, not tests.
Raj's account is a single user's experience. LM Studio's documentation supports the mechanism and its boundaries, and Google's Gemma documentation supports the memory reasoning. Nobody has yet published independent speed or reliability numbers for this specific combination.