Local AI on the iPhone finally becomes useful beyond the chat window

Faster models, better integrations and tools such as Siri and Shortcuts are turning on-device AI from a novelty into a practical, private utility.

Local AI on the iPhone has moved from a curiosity to something genuinely useful, but only because the software has caught up with the models. A year ago, running a tiny language model on a handset felt like a novelty: responses were slow, the answers were brittle and cloud-based assistants remained the obvious choice. That has changed. Current small models such as Qwen, Gemma and Liquid AI’s LFM family are now compact enough to run on-device, while Apple’s Foundation Models framework is also opening the door to third-party apps that can tap into local processing more directly.

What makes the new wave more compelling is not just the models, but the way apps expose them. In practice, the best tools now sit inside the places people already work: Siri, Shortcuts, the Action Button and Control Centre. That matters more than a polished chat window. It means local AI can be used for real tasks such as transcribing voice notes, cleaning up draft text, summarising documents and handling private prompts without sending data to a server first.

One of the most flexible options is Locally, now LM Studio’s official iPhone app after the desktop company acquired the original indie project earlier this year. The app runs models through Apple’s MLX framework and requires no account. Its model list includes Qwen, LFM 2.5, Gemma, Llama and Apple’s own Foundation model, with image input available on some variants. The app also supports voice access through quick actions and shortcuts, which makes it easier to use as a utility rather than a chat destination. A more significant feature is LM Link, released in June, which lets the phone act as a secure front end to LM Studio running on a Mac at home. In that mode, the heavier computation moves off the handset, so users can work with larger models without the same battery and thermal constraints.

Private LLM takes a different approach by focusing on automation. It is a one-off purchase rather than a subscription, and it integrates with Siri and Shortcuts so an AI model can become one step in a workflow instead of a separate app session. The developer says it uses MLC-LLM with OmniQuant quantisation, which is intended to preserve output quality better than simpler compression schemes. The app is limited to text, but that restriction suits use cases such as dictation cleanup, editing and summarisation. In testing, a Shortcut that rewrote spoken notes into clean prose was accurate enough to paste directly into an email, and even a long article was handled without the app freezing.

PocketPal AI is the most open-ended of the group. It is free and open source, built on llama.cpp, and it lets users browse Hugging Face from inside the app, select any GGUF model and choose a quantisation level that matches available memory. That makes it useful for experimentation, particularly when a new compact model appears and can be downloaded the same day. It also supports saved assistants called Pals and a community hub for shared prompts. Because it supports iOS 15.1 and later, it is also the broadest compatibility option here, which is important for older devices that cannot run the latest software.

Google AI Edge Gallery is less of a daily assistant and more of a development playground. It is aimed at Google’s Gemma models and adds features such as Thinking Mode, which displays a model’s reasoning process, and Agent Skills, which can supply tools such as Wikipedia look-ups, maps or visual summary cards. The app is still marked as being in active development, and that roughness shows, but it is a useful glimpse of where local AI interfaces may be heading: less like static chat and more like modular software with built-in tools. Enclave AI follows a more document-focused path, with local voice chat, PDF handling, Siri support and a Shortcuts action, while also offering access to cloud models such as Claude and Gemini for users who want the option of switching between local and remote processing.

The practical lesson from all of this is that choosing a model matters less than many newcomers expect, provided the device can actually run it. The key numbers are the parameter count, usually shown with a “B”, and the quantisation format, often shown in file names such as Q4_K_M or Q8_0. Smaller parameter counts are faster and easier on memory; heavier models generally produce richer answers but take longer and consume more storage. Quantisation is essentially compression: lower-bit files are smaller, while higher-bit files preserve more of the original model. On an iPhone, the safest starting point is usually a model in the 2B to 4B range, with Q4_K_M often being the most practical file type. Storage can also become a hidden cost, because each app typically stores its own copy of the same model. In other words, the real constraint is not whether local AI works on a phone; it is how much room the phone has to spare, and whether the app makes the model useful enough to justify keeping it installed.