Pocket AI Lab
  1. Home
  2. Models

Models that run offline on iPhone

The curated Pocket AI Lab catalog has 23 models across five backends, from an 80 MB Whisper to a 14B Qwen. 18 of them have measured results on an iPhone Air. Not sure which to pick? See the best local LLM for your iPhone.

Catalog

23 models

⚡ is generation speed on iPhone Air in app 1.1.0. * means the phone throttled during that run.

How much memory does each model need?

From 500 MB for Whisper tiny to 8.4 GB for Gemma 3 12B, as budgeted in the catalog. Peak is what the app actually measured on the phone.

ModelQuantRAM GBDisk GBPeak MB
Whisper tiny (English)whisper.cppggml0.50.08–
Whisper basewhisper.cppggml0.80.15–
LFM2 1.2B (GGUF)llama.cppQ4_K_M1.40.80249
Qwen 3.5 0.8B (GGUF, Vision)llama.cppQ4_K_M1.50.80598
Gemma 2 2BMLX4bit1.61.502,148
Llama 3.2 3BMLX4bit2.22.002,568
Qwen 2.5 3BMLX4bit2.22.002,360
LFM2 2.6B (GGUF)llama.cppQ4_K_M2.21.60266
Phi-3.5 MiniMLX4bit2.42.303,189
Llama 3.2 3B (GGUF)llama.cppQ4_K_M2.52.10721
Qwen2.5 1.5B (Core ML)Core MLint8-W8A162.52.304,354
Phi-3.5 Mini (GGUF)llama.cppQ4_K_M2.72.401,739
Qwen 3.5 2B (GGUF, Vision)llama.cppQ4_K_M2.82.001,054
Qwen 3.5 4B (GGUF, Vision)llama.cppQ4_K_M4.43.501,252
Mistral 7B v0.3MLX4bit4.54.404,671
Llama 3.1 8BMLX4bit4.74.905,121
Gemma 3 4B (Vision)MLX4bit4.82.804,577
Gemma 3 4B (GGUF, Vision)llama.cppQ4_K_M5.03.402,175
Gemma 4 E2B (GGUF, Vision)llama.cppQ4_K_M5.04.101,490
Qwen2-VL 7B (Vision)MLX4bit5.24.705,461
Qwen 2.5 14BMLX4bit8.28.50–
Gemma 3 12B (Vision)MLX4bit8.47.50–

RAM and disk come from catalog.json. Peak MB is the measured process footprint (phys_footprint) during the benchmark. llama.cpp maps weights from disk, so its footprint stays far below the file size.

How does the app decide what fits my iPhone?

It compares each model's memory need with two limits it reads from your phone at launch: the per-app allowance iOS grants before it kills the app, and the GPU working set.

The phone's RAM figure isn't what an app gets. iOS kills a process that crosses its memory limit (jetsam), and that limit sits well below physical RAM. Pocket AI Lab requests Apple's increased-memory-limit entitlement and then measures the real allowance with os_proc_available_memory.

MLX and Core ML copy weights into memory, so they're checked against that allowance. llama.cpp and whisper.cpp map the file instead, so they're checked against the larger Metal working set. The model offered as recommended has to stay under 85% of its budget, because a model that barely fits loads and then dies once the context grows.

Can I add a model that isn't in the catalog?

Yes. In the Models tab tap ⊕ and paste a Hugging Face repo, or pick a GGUF file from the Files app.

The catalog itself is a JSON file in the GitHub repository. The app refreshes it on launch, so new curated models show up without an App Store update.