v1.1.0iOS 26.1+MITno account
Pocket AI Lab Offline LLM app for iPhone
Pocket AI Lab runs open language models from Hugging Face entirely on your iPhone. It checks how much memory iOS will actually give the app, tells you which models fit, downloads them once and then works in airplane mode. There's no account, and your conversations stay on the phone. Wondering which model to start with? See the best local LLM for your iPhone.
- Free
- 41.1 MB
- Any iPhone on iOS 26.1
- English
- 5inference backends in one app
- 23models in the curated catalog, plus any GGUF you import
- 77.3 tok/sLFM2 1.2B (GGUF) (llama.cpp) on iPhone Air, fastest measured
- 0accounts. Chats, attachments and voice recordings stay on the phone
What is Pocket AI Lab?
Pocket AI Lab is a free iPhone app that downloads open LLMs from Hugging Face and runs them on the phone's own chip, so chats work without internet and never leave the device.
It's built like a small lab bench. You see which runtime is generating, how fast, how hot the phone is getting and how much memory the model really uses. The same model can be run on different backends and measured with the built-in benchmark, which grades it on 118 tasks and produces a report card you can share.
Who is it for?
It's for people who want a private assistant that works offline, and for developers who want to know how a model actually behaves on an iPhone before they build on it.
Private, offline chat
On a plane, in the metro, or just when you'd rather not paste things into a cloud chatbot. After the download, the model runs with no network at all.
Picking an on-device model
Compare GGUF against MLX, check tokens per second, load time and peak memory on your own phone, and export the report as JSON.
Trying new models
Paste a Hugging Face link, pick a quantization and get a fits / doesn't fit verdict for your device before a single byte downloads.
How does it work?
The app scans the phone, filters the catalog against the memory iOS really allows, downloads the model from Hugging Face and streams the answer token by token from whichever backend fits the model.
Scan
Chip, RAM, free disk, the per-app memory allowance from
os_proc_available_memoryand the Metal working-set limit.Recommend
MLX and Core ML models are judged against the strict app limit. Memory-mapped llama.cpp and whisper.cpp models against the higher GPU working set.
Download
Background download pinned to an exact repo revision. It survives app restarts, and the catalog itself updates from GitHub.
Chat
Every backend speaks one streaming interface, with full conversation history and live CPU, RAM and thermal readings.
Which inference backends does it support?
Five: llama.cpp for GGUF files, MLX Swift for MLX weights and vision models, Core ML, Apple's built-in Foundation Models, and whisper.cpp for speech-to-text.
llama.cpp
Runs GGUF files, the most common format for quantized models on Hugging Face. Weights are memory-mapped, so a 3B model at Q4_K_M peaked at 721 MB in our sweep. Vision models work through a separate mmproj projector.
MLX Swift
Apple's array framework for Apple silicon. Runs the 4-bit models from mlx-community, plus vision-language models that take photos and video frames.
Core ML
Compiled Core ML packages, loaded through swift-transformers. On the current catalog model prompt reading is fast (about 646 tok/s) and generation slow (3.0 tok/s).
Apple Foundation Models
The ~3B model that ships with Apple Intelligence. Nothing to download. Needs an Apple Intelligence iPhone with the feature turned on.
whisper.cpp
Speech-to-text. Record a voice message and the Whisper model turns it into text on the phone. 80 MB for tiny.en.
Import by link
Paste owner/repo from Hugging Face. GGUF, MLX safetensors, Core ML and Whisper files are detected, Q4_K_M is preselected.
How fast do LLMs run on an iPhone?
On an iPhone Air with the A19 chip, 2B to 4B models generated 26.3 to 36.6 tokens per second in Pocket AI Lab 1.1.0. That's faster than most people read.
| Model | Gen tok/s | pp512 / tg128 | Load s | Peak MB | Score |
|---|---|---|---|---|---|
| Qwen 3.5 2B (GGUF, Vision) | 36.6 ± 1.3 | ~854 / 33.1 | 0.4 | 1,054 | 64.4 |
| LFM2 2.6B (GGUF) | 35.9 ± 0.7 | ~560 / 33.1 | 0.1 | 266 | 60.2 |
| Gemma 2 2B | 30.3 ± 0.3 | ~277 / 27.4 | 3.7 | 2,148 | 40.7 |
| Llama 3.2 3B (GGUF) | 26.9 ± 0.6 | ~459 / 21.9 | 0.2 | 721 | 59.3 |
| Qwen 2.5 3B | 26.3 ± 2.8 | ~249 / 22.4 | 1.7 | 2,360 | 55.1 |
Measured 2026-09-03 to 2026-10-23 on iPhone Air (A19, 11.5 GB RAM as reported by iOS, iOS 26.6.1). Gen tok/s is the mean of 3 runs of 200 tokens. Full table and method on the benchmarks page.
Can I run any model from Hugging Face?
Mostly yes. Paste a public repo link and the app reads the file list, offers the available quantizations and shows a RAM and disk verdict for your phone before downloading.
GGUF files from the Files app work too. Gated repos (the ones where you accept a license on the website first) can't be downloaded, because the app downloads anonymously and never asks for your Hugging Face token. A mirror repo of the same model usually works.
Models → ⊕ → paste
bartowski/Llama-3.2-1B-Instruct-GGUF
The app validates the repo through the Hugging Face API, detects the format, then compares the file size with your phone's memory budget. Illustration of the import flow, not a screenshot.
What do I need to run it?
Any iPhone on iOS 26.1 or later. How big a model it runs depends on its memory: 1B models on 4 GB iPhones, up to 2–3B on 6 GB, up to 4B and Gemma 4 E2B on 8 GB, and the larger vision models on 12 GB.
- Apple Foundation Models also need Apple Intelligence turned on, so iPhone 15 Pro or newer.
- Storage depends on the model, from 80 MB for Whisper tiny to 8.5 GB for Qwen 2.5 14B. The app itself is 41.1 MB.
- The app is universal, so it installs on iPads with iPadOS 26.1 too. Published benchmarks so far come from an iPhone.
Is Pocket AI Lab open source?
Yes. The full Swift source is on GitHub under the MIT license, and the model catalog the app downloads lives in the same repository.