Local AI on iPhone finally becomes useful beyond the novelty

A new generation of iPhone apps makes offline AI practical, linking local models to Siri, Shortcuts, documents and even desktop hardware while keeping more tasks on the device.

Running a local AI model on a phone has shifted from novelty to something genuinely useful. The central change is not only in model quality, but in the apps built around them. After an early experiment with a tiny model in 2024 ended in frustration, the recent generation of iPhone apps now ties on-device models into Siri, Shortcuts, the Action Button and even remote desktop access, making them far easier to reach in everyday use. That matters most when a cloud service is unavailable, or when the task is better kept on the device. According to the MakeUseOf article that inspired this piece, the author tested the latest options on an iPhone 17 Pro Max running iOS 27 and found several strong contenders.

Locally, now the official iPhone app from LM Studio, is the clearest example of how far the category has come. The App Store listing says the app runs offline, requires no login and does not collect data, while supporting models such as Llama, Gemma and Qwen. The company also added LM Link, which lets the phone connect securely to a desktop machine running LM Studio, with end-to-end encryption and chat history stored locally on each device. In practice, that means a smaller model can run directly on the handset, while a larger one can be accessed remotely when more capable hardware is available. The MakeUseOf reviewer found the phone-based Qwen 3.5 response quick enough for simple prompts and noted that the remote setup delivered more expansive answers, albeit with a slight delay.

Private LLM is aimed less at casual chatting and more at automation. The app costs $4.99 once, supports Family Sharing and integrates with Siri and Shortcuts, so it can be inserted into workflows rather than opened as a separate chat interface. The developer says it uses the MLC-LLM engine with OmniQuant quantisation, a method intended to preserve quality while reducing the memory footprint of the model. The app is text-only, but that limitation did not prevent it from handling useful tasks. In the MakeUseOf tests, it cleaned up dictated notes well enough to paste directly into email and summarised a long article without freezing.

PocketPal AI stands out for flexibility. It is free and open source, built on llama.cpp, and can search Hugging Face from inside the app so users can download almost any GGUF model they want, then select the quantisation level that best matches the phone’s memory. The app also includes saved assistants called Pals and a community hub for sharing them. Because it supports iOS 15.1 and later, it is also the most accommodating option for older devices. The reviewer used it to download a Gemma 4 model from Hugging Face and received a detailed explanation of model parameter sizes within seconds.

Google AI Edge Gallery is less polished as a daily assistant, but it is valuable as a test environment for Gemma models. It is free, requires iOS 17 or later and includes Thinking Mode, which exposes the model’s reasoning process in a visible form. More notably, it offers Agent Skills, which add tools such as Wikipedia look-ups, interactive maps and visual summary cards, with the option to load more skills from a URL. Google says the app remains under active development, and that roughness shows, but the feature set is unusually ambitious. In the review, the Wikipedia skill worked reliably for fact-checking questions, while Thinking Mode produced broadly similar results without a major speed penalty.

Enclave AI is the most practical choice for documents and offline voice interaction. It is free with in-app purchases for larger models, runs on llama.cpp and supports local voice chat even without an internet connection. It also adds direct document handling, along with Siri support and an “Ask Enclave” Shortcut. The app can connect to cloud models such as Claude and Gemini, although that makes it important to check which model is active before sharing sensitive material. The author found document answers accurate, if somewhat slower, and noted that the offline voice mode worked, albeit with robotic audio and noticeable lag.

For newcomers, the biggest decisions are not the app names but the model size, the quantisation level and the available memory. In practical terms, the parameter count gives a rough sense of capability: smaller models are faster and less demanding, while larger ones usually write better and know more. Quantisation is a compression trade-off, and Q4_K_M is often the safest starting point on a phone because it balances size and quality. Storage is also easy to underestimate, especially because each app keeps its own copy of a model. The broader lesson from these new iPhone tools is that local AI is no longer just for enthusiasts. It is becoming a useful fallback for private notes, quick edits and offline tasks, even if larger cloud models still remain better for serious research.