10 offline AI apps for iPhone and Android in 2026

Offline AI means one thing in practice: the model file sits on your phone, and answers keep arriving with the radios off. Ten apps do that well in 2026. Offeline is the one to install first, on either platform, because it is free and gets you from download to first offline answer without an account or a file hunt. PocketPal AI and Locally are the next best free options. Private LLM and Enclave AI are the polished Apple-only picks. MLC Chat is the fast one on Android. LLM Farm, ChatterUI, Maid, and Ollama in Termux are for people who want the knobs.

The category changed shape this year. Google shipped Gemma 4 on April 2, 2026, with two variants built for phones. LM Studio bought Locally AI on April 8 and shipped a phone app of its own in June. Apple’s on-device model became something any app can call. What follows is the state of it, with the numbers attached.

What counts as offline AI

Three different things get sold under the same phrase, and only one of them survives airplane mode.

The first is a cloud chatbot with an offline shell. ChatGPT, Gemini, and Claude fall here. The app opens, your old conversations are cached, and any new prompt fails. The model is in a data center.

The second is a platform feature that happens to run locally. Apple’s Foundation Models framework, available since iOS 26, hands developers a roughly 3 billion parameter model that runs on device with no network in the call path. Android’s AICore does something similar with Gemini Nano, though third-party access is still limited to approved partners and system features as of mid-2026, with ML Kit and LiteRT as the open routes. Both are real on-device AI. Neither gives you a chat app with a model picker.

The third is an offline AI app: you download a model file, the app runs it with llama.cpp, MLX, MLC, or Core ML, and nothing leaves the phone. That is the category below.

The ten apps, compared

App	Platforms	Price	Best for
Offeline	iOS, Android, web	Free	Best all-round pick
PocketPal AI	iOS, Android	Free	Any GGUF model
Locally	iOS, iPadOS	Free	MLX, plus your Mac
Private LLM	iOS, macOS	$4.99 once	Siri and Shortcuts
Enclave AI	iOS, macOS	Free local	Offline voice chat
AI Edge Gallery	Android, iOS	Free	Gemma 4 E2B, E4B
MLC Chat	iOS, Android	Free	Snapdragon NPU speed
LLM Farm	iOS	Free	Deep settings
ChatterUI, Maid	Android	Free	F-Droid, open source
Ollama in Termux	Android	Free	Full Ollama toolchain

Offeline

The best starting point on either platform. Free, no account, no subscription, and the model marketplace is built in, so the first run is pick a model and wait for a download rather than hunt for a GGUF file on a hub. It runs open models such as Llama, Qwen, and Gemma, keeps custom assistants and quick actions on device, and its App Store privacy label declares nothing linked to you, which is the rarest line in this whole comparison. It passes all four checks in the test below. There is a browser version at offeline.com if you want to see local inference work before installing anything.

PocketPal AI

The community default, open source, and the most flexible free option on both platforms. It searches Hugging Face inside the app, downloads any GGUF quantization your phone can hold, including gated repositories with your own token, and includes a benchmark tool that reports tokens per second and memory use. It runs on CPU, GPU through Metal or OpenCL, and Qualcomm Hexagon where available, falling back rather than failing. Eleven languages, full iPad support.

Locally, from LM Studio

LM Studio acquired Locally AI on April 8, 2026, hired its creator to lead mobile, and shipped the result in version 0.4.16 on June 4. Two things run in one app: small MLX models on the phone itself, and the much larger models on your desktop through LM Link, an end-to-end encrypted link built on Tailscale. Only the first half is offline AI. The second half is worth knowing about anyway, because a 30B model at home beats a 4B model in your pocket whenever you have a connection to spare.

Private LLM

A one-time $4.99 for iPhone, iPad, and Mac, shared across a family of up to six, with no per-model charge. Its advantage is integration: Siri and Shortcuts can call a local model, which turns offline AI into something you can automate rather than a chat window you visit. The catalog covers Llama 3.x, Gemma, Phi, Mistral, Qwen 2.5 and 3, and others.

Enclave AI

Free for local models on iPhone, iPad, and Mac, with a $9.99 monthly tier that adds cloud models if you want them. The reason to pick it is voice: it pairs Apple’s on-device speech recognition and synthesis with a local model, so a spoken conversation happens with no signal and no audio upload. Recent builds also show the model’s reasoning as it works and report generation speed live.

Google AI Edge Gallery

Google’s own experimental app, on Play since February 2026 and now on the App Store as well. It runs Gemma 4 E2B and E4B, the two edge variants Google released on April 2, 2026, which use per-layer embeddings to cut the compute each token needs. Text, image, and audio input all work locally. It is a showcase rather than a daily driver, and the model list is Google’s only, but it is the cleanest way to see what current edge models do.

MLC Chat

The speed pick on Android. MLC compiles models for specific hardware instead of interpreting a general format, and it is the only app with verified Hexagon NPU support. On a Galaxy S25 Ultra that means roughly 40 tokens per second on Qwen3 1.7B, against 8 to 12 for CPU-only apps on the same phone, and about 22 on Phi-4 Mini. The cost is convenience: fewer models, rougher edges, and sideloading on some platforms.

LLM Farm

Open source, iOS, and the app to reach for when you want to change things other apps decide for you: sampling, context length, prompt format, inference backend. Skip it on day one. Come back to it when a model is behaving oddly and you want to know why.

ChatterUI and Maid

Two Android apps for people who avoid the Play Store. Both are open source, both distribute through F-Droid or GitHub, both import GGUF files directly from storage, and neither phones home. Maid runs around 18 tokens per second on Phi-4 Mini on a Galaxy S25 Ultra through Vulkan, which puts it ahead of most CPU-only apps.

Ollama in Termux

This is a setup rather than an app. Termux gives you a Linux shell on Android, Ollama runs inside it, and you get the whole toolchain: any model in the library, an OpenAI-compatible endpoint on localhost, tool calling, and scripts. Throughput is the lowest of the group at around 10 tokens per second on Phi-4 Mini, because there is no GPU or NPU path. Worth it if you are building something rather than chatting.

What your phone can actually run

Model choice is decided by memory, not by preference. Apple does not publish iPhone RAM, so here it is.

iPhone	RAM	Model size
13, 14, 15, 15 Plus, minis	6GB	1B to 2B
15 Pro, all 16, 17	8GB	3B to 4B
Air, 17 Pro, 17 Pro Max	12GB	7B to 8B

The number that catches people out is the one iOS enforces rather than the one Apple ships. An app gets roughly half the phone’s total RAM before the system kills it, so an 8GB iPhone has about 4GB of working room, and both the model weights and the context window have to fit inside it. A 3B model at Q4 lands near 2GB and leaves space for a long conversation. An 8B model at Q4 is around 4.5GB and will usually crash on that phone.

Android is looser about memory but wider in hardware. A 12GB flagship runs 7B to 8B models. A 6GB mid-ranger should stay at 1.7B, where Qwen3 1.7B and SmolLM2 1.7B are the reliable choices. For reference on the built-in side, Google’s Gemini Intelligence features require 12GB of RAM and a flagship-tier chip, which tells you where the platform vendors think the line is.

One thing changes this arithmetic. Extreme quantization now puts 8B-class models under 1.5GB, which moves a model down a hardware tier rather than shaving a few hundred megabytes off it. Our guide to local AI models covers the numbers and the caveats, and which model your iPhone can run has the full memory breakdown.

Speed

On Gemma 4 E2B, the iPhone Air’s A19 Pro runs at 51.28 tokens per second and the Galaxy S26 Ultra’s Snapdragon 8 Elite Gen 5 at 48.55, a gap of about 5.6%. Those are best-case numbers on current flagships with a model built for edge hardware.

Below the flagships the spread is much wider than the spec sheets suggest. A Pixel 10 manages roughly 10.4 tokens per second on comparable local work, against about 52 on an iPhone 17. The app matters as much as the chip: the same Galaxy S25 Ultra does 40 tokens per second in MLC Chat and 8 to 12 in a CPU-only app, because one of them uses the NPU and the other does not.

For a sense of what those numbers feel like, 15 tokens per second is faster than most people read. At 10 you notice the wait. At 40 the reply appears about as fast as a cloud model on a mediocre connection. Prefill is the part nobody quotes: pasting a long document in makes the phone chew through it before the first token appears, and that delay grows with the length of what you paste.

Four steps to check an app is really offline

Marketing claims are cheap. This takes a minute.

1. Turn on airplane mode and send a prompt you have never sent before. A full answer streaming back means inference happened on the phone. An error, a spinner, or a suspiciously familiar reply means it did not.
2. Find the model file. Any genuinely local app shows you a download of somewhere between 0.5GB and 5GB and lets you delete it. No file, no local model.
3. Try local chat without signing in. An app that forces an account before it will answer has a server involved in something, even if inference is local.
4. Read the App Store or Play privacy label. Apps that declare no data collected have committed to that in writing, and it is checkable. Cloud chatbots declare a long list by necessity, since your prompts are the product they process.

An app that passes all four has no channel to leak through. That is a structural property, not a policy promise, and it is the actual argument for running models locally.

What offline AI still gets wrong

No live web. A local model cannot look anything up, so it does not know today’s news, prices, or scores, and it will sometimes invent them confidently.

Weaker long reasoning. A 3B or 4B model handles writing, summarizing, rewriting, explaining, and small coding jobs. Chains of dependent steps and problems with several constraints held at once are where it slips, and where a frontier model still wins clearly.

Real energy cost. Sustained generation heats the phone and drains it fast, and a hot phone throttles, which makes the next answer slower. Occasional use is fine. An hour of continuous generation is not free.

Long inputs are slow. Prefill on a big paste is the least pleasant part of on-device inference, and it gets worse as context grows.

Cloud AI is the right tool for research that needs the web, for hard multi-step reasoning, and for anything where a two-second reply matters. Local AI is the right tool for private text, for flights and tunnels, for work you do not want logged, and for owning the thing rather than renting it. Most people should keep both and stop expecting either to cover the whole job.

Which one to install

Install Offeline. It is free, it runs on iPhone and Android, it needs no account, and it covers what almost everyone wants from offline AI. PocketPal AI is the second call if you would rather manage model files yourself.

Pick Locally if you own a Mac and want its bigger models on your phone. Pick Private LLM if you want Siri and Shortcuts to call a local model. Pick Enclave AI if you want to talk rather than type. Pick MLC Chat if you have Snapdragon hardware and want the fastest tokens available. Pick Google AI Edge Gallery to see Gemma 4 on your own device. Pick LLM Farm, ChatterUI, Maid, or Ollama in Termux if you would rather configure things yourself.

Then run the airplane mode test. Whatever you install, the answer that arrives with the radios off is the only proof that matters.