- Home
- FAQ
Pocket AI Lab FAQ
Short answers about the offline LLM app for iPhone: privacy, supported iPhones, models, storage and speed.
Does Pocket AI Lab work offline?
Yes. Once a model is downloaded, chat, vision and speech-to-text run on the iPhone with no network connection, including in airplane mode.
You need internet only to download models from Hugging Face and to refresh the model catalog.
Does Pocket AI Lab send my chats anywhere?
No. Conversations are processed and stored on the device: prompts, attachments and voice recordings are handled by the model running on your iPhone and never leave the phone.
The app only talks to huggingface.co (model files), raw.githubusercontent.com (catalog updates), itunes.apple.com (a version check) and ingest.proapp.uk (how models perform on the device). Details on the privacy page.
Which iPhones can run local LLMs in Pocket AI Lab?
Any iPhone on iOS 26.1 or later, from iPhone 11 and SE onwards. Memory decides how big a model it runs: 1B models such as LFM2 1.2B on 4 GB iPhones, 2–3B models such as LFM2 2.6B or Llama 3.2 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 need Apple Intelligence, which means iPhone 15 Pro or newer with the feature turned on. Bigger models like Qwen 2.5 14B need more memory than most iPhones give a single app, and the app will tell you before you download.
How much storage do the models take?
From 80 MB for Whisper tiny (English) to 8.5 GB for Qwen 2.5 14B. A typical 3B chat model at 4-bit is about 2 GB.
The app itself is a 41.1 MB download. Before a download starts, the app checks free disk space so it fails early instead of halfway.
Which models can I run?
The curated catalog has 23 models: Llama 3.1 and 3.2, Gemma 2 and 3 (with vision), Qwen 2.5 and Qwen2-VL, Phi-3.5 Mini, Mistral 7B, Whisper, and Apple's built-in model.
You can also import public models from Hugging Face by link (GGUF, MLX, Core ML and Whisper files) or open a GGUF file from the Files app. See all models.
How fast is it?
On an iPhone Air, 2B to 4B models generated roughly 26.3 to 36.6 tokens per second, and 7B to 8B models around 8 to 12 once the phone heated up.
Every number with its method is on the benchmarks page.
Is Pocket AI Lab free?
Yes. It's free on the App Store, with no subscription and no in-app purchases in the current version.
Models are free downloads from Hugging Face, each under its own license.
Is Pocket AI Lab open source?
Yes. The Swift source code is on GitHub under the MIT license.
It lives at github.com/ananasDDA/pocket-ai-lab, together with the model catalog the app downloads.
Should I pick the GGUF or the MLX version of a model?
Try GGUF first if memory is tight. In our tests Llama 3.2 3B on llama.cpp peaked at 721 MB against 2,568 MB on MLX, at a similar speed.
MLX is the route to most vision-language models in the catalog, and it's worth benchmarking both on your own phone.
Why does the app say a model won't fit when my iPhone has 8 GB of RAM?
Because iOS doesn't give one app all of the RAM. It kills apps that pass a per-app limit, and that limit is well below the number on the box.
Pocket AI Lab reads the real allowance on your phone and checks each model against it, so you don't download 5 GB just to watch the app close.
Does it work on iPad?
The App Store build is universal, so it installs on iPads with iPadOS 26.1 or later.
Benchmarks published so far come from an iPhone.
How is Pocket AI Lab different from PocketPal or Private LLM?
Pocket AI Lab runs several backends side by side and measures models on your phone, PocketPal focuses on GGUF via llama.cpp and also runs on Android, and Private LLM is a paid app with a curated model list and Siri support.
The full, sourced comparison is on the compare page.