To run an LLM on iPhone, check the model file and app runtime first. This guide gives a procedure for model selection. It does not claim a measured Q4-versus-Q8 result or recommend unverified iPhone model files.
We checked the linked vendor sources on October 14, 2026. This post reports no comparative product test.
1. Identify the exact model
Record the model name, version, source, and license. Do not select a file from its parameter count alone.
2. Confirm the iPhone app runtime
Check the app documentation for the model format and architecture. Confirm support for the exact file before download.
PocketLLM is pre-launch. Private, on-device AI, starting on iPhone, with more platforms planned. No account required, no in-app tracking, and chat processing stays on your device. Join the waitlist.
3. Read the llama.cpp quantization label
The llama.cpp quantization guide documents formats such as Q4_K_M and Q8_0. A format label does not establish task accuracy.
4. Check the file source
Use the original model card and the conversion record. Check that the file corresponds to the model you selected.
5. Check memory on your device
Use the context length that your task needs. Record the device, app version, and model file with any memory result.
6. Compare the same task
Keep the prompt and generation settings the same between variants. Change one variable at a time.
7. Check output quality
Verify facts, calculations, and the required format. Do not select a variant only because its file is smaller.
8. Check sustained use
Repeat a realistic task over the session length you need. Observe heat, battery use, and app stability on your device.
9. Keep a reproducible record
Save the exact configuration and example outputs. State the test conditions before you publish a performance comparison.
PocketLLM uses a separate catalog
Its current catalog includes Llama 3.2 1B and Phi-4 Mini. See the model catalog for the app entries. The recorded evaluation does not establish a matched Q4-versus-Q8 comparison.
Local inference keeps prompts on the device during inference when the selected model runs there. For a local-only workflow, select no cloud model. Disable chat sync and cloud backup. Check the app telemetry policy. Disable optional network features. Review extensions, exports, and local storage access.
Frequently asked questions
Which quantization formats does llama.cpp document?
The llama.cpp quantization guide documents formats such as Q4_K_M and Q8_0.
Does PocketLLM list Llama 3.2 1B and Phi-4 Mini?
The current PocketLLM model catalog includes Llama 3.2 1B and Phi-4 Mini.
Does the PocketLLM evaluation establish a matched Q4-versus-Q8 comparison?
The recorded PocketLLM evaluation does not establish a matched Q4-versus-Q8 comparison.
PocketLLM is pre-launch on iPhone, with more platforms planned.