PocketLLM runs single-file GGUF models on the iPhone with llama.cpp and Apple Metal. Whether a model runs depends on its architecture, quantization, context size and the memory iOS allows the app; PocketLLM checks what it can before loading and explains any failure.
These models are listed under Models → Recommended and download directly from Hugging Face. All are Apache-2.0 or MIT licensed and need no account.
| Model | Size (Q4_K_M) | Licence | Notes |
|---|---|---|---|
| Qwen3 4B Instruct 2507 | 2.3 GB | Apache-2.0 | Tested on iPhone 17 Pro: chat, streaming, tool calling |
| Qwen2.5 0.5B Instruct | 0.5 GB | Apache-2.0 | Tested: chat, streaming, tool calling. Smallest and fastest |
| SmolLM2 1.7B Instruct | 1.0 GB | Apache-2.0 | |
| Qwen2.5 1.5B Instruct | 1.0 GB | Apache-2.0 | Multilingual |
| Qwen3 1.7B | 1.2 GB | Apache-2.0 | Reasoning |
| Phi-3.5 Mini Instruct | 2.2 GB | MIT | |
| Phi-4 Mini Instruct | 2.3 GB | MIT | Reasoning |
| Mistral 7B Instruct v0.3 | 4.1 GB | Apache-2.0 | Needs a recent iPhone with more memory |
| Qwen2.5 7B Instruct | 4.4 GB | Apache-2.0 | Needs a recent iPhone with more memory |
| Qwen3 8B | 4.7 GB | Apache-2.0 | Needs a recent iPhone with more memory |
Expand any model in the app to see whether it fits your iPhone’s memory right now (an estimate, shown as such).
Tap + in the Model Library to import any .gguf file from Files.
model-00001-of-00002.gguf) must be
merged first.