Which AI model should you run on your iPhone? A guide to local LLMs
· By Modelbase Team
Modelbase has more than 50 open models, and the right one depends on two things: what you want to do, and how much memory your device has. Here’s how to choose.
The rule of thumb
A model has to fit in your device’s memory while it runs, next to the system and your other apps. Its download size is a good guide to how much it needs: a model of 1 GB needs about 1 GB of memory, and a model of 3 GB about 3 GB.
- iPhones with 8 GB of memory (iPhone 15 Pro and later): models up to 4B are a good fit.
- iPhones with 6 GB or less: models up to about 2B are the better choice.
- iPad and Mac: most have more memory than a phone, so the larger models are a good place to start.
Bigger models know more and write better. Smaller ones answer faster and leave room for other apps. You don’t have to work it out alone: every model’s card in Manage models says which iPhones it’s recommended for, such as “iPhone 15 and newer” or “iPhone 15 Pro and newer”.
Start here
LFM 2.5 (1.2B) comes with the app, at 0.66 GB. It’s quick on recent iPhones and handles everyday writing, questions and summaries well. Many people never need anything else. On an iPhone 14 or older, LFM 2.5 (350M) (0.23 GB) is the lighter model the app recommends.
Pick by what you need
| You want | Try | Size |
|---|---|---|
| The lightest models, for older iPhones | LFM 2.5 (350M), Granite 4.0 H (350M), Gemma 3 (270M) | 0.19–0.23 GB |
| Speed and a small download | Qwen 3.5 (0.8B), Gemma 3 (1B), Llama 3.2 (1B) | 0.65–0.77 GB |
| A good all-rounder | Qwen 3.5 (2B), Llama 3.2 (3B), Granite 4.0 H Micro | 1.75–1.82 GB |
| The best answers on a phone | Qwen 3.5 (4B), Phi 4 Mini (3.8B), Gemma 3 QAT (4B) | 2.18–3.06 GB |
| Questions about photos | LFM 2.5 VL (1.6B), Qwen 3 VL (2B), Gemma 4 (E2B) | 1.5–3.58 GB |
| Step-by-step reasoning | LFM 2.5 Thinking (1.2B), Qwen 3 Thinking (4B), DeepSeek R1 (1.5B) | 0.66–2.28 GB |
Thinking, or not
Some models can reason before they answer. The picker next to the message field sets how much: Instant answers straight away, while Medium and High take longer and do better on math, logic and planning. Everyday questions rarely need more than Instant. Models that always reason, such as LFM 2.5 Thinking (1.2B) and DeepSeek R1 (1.5B), offer Medium and High only. What is a reasoning model? explains what each level does.
How to switch models
Tap the model’s name at the top of a chat to open Select model: the models on your device are listed under “On device”. To add one, tap Manage models, then Local, find a model by name, maker or size, and tap Download. When the card shows a checkmark, pick the model in Select model. For a closer look at single models, see the guides to LFM 2.5, Llama 3.2 and Phi 4 Mini.
When a phone isn’t enough
For the largest models, connect OpenRouter with your own key, or add a server running on your own computer, such as LM Studio or Ollama. Its models appear in the picker next to the ones on your device, and you can switch between them in any chat. Some cloud APIs have free tiers; see free LLM APIs.
Torn between two families? See Gemma vs Llama vs Qwen and Qwen vs DeepSeek.
You can download, try and delete models as often as you like in Manage models: nothing is lost by experimenting.
Try it on your iPhone, iPad or Mac
Modelbase is free on the App Store.