Choosing a model
Switch models in Settings → Model. A change applies to your very next message — settings are read per turn rather than cached, so you do not need to restart a conversation.
The catalogue
Downloadable entries come from the mlx-community organisation on Hugging
Face, which hosts pre-converted weights. Apple Intelligence is built into iOS
and downloads nothing.
| Model | Size | Needs | Notes |
|---|---|---|---|
| Apple Intelligence | — | iOS 26 | Built in, starts instantly |
| Llama 3.2 1B Instruct | ~0.7 GB | 4 GB | Default. Fastest; any supported iPhone |
| Qwen 3 1.7B | ~1.0 GB | 4 GB | Noticeably stronger than Llama 1B for the size |
| Llama 3.2 3B Instruct | ~1.8 GB | 6 GB | Solid general purpose |
| Qwen 3 4B | ~2.3 GB | 6 GB | Best all-round that still fits a 6 GB iPhone |
| Gemma 3n E2B | ~2.5 GB | 6 GB | ~5B parameters behaving like a 2B |
| Phi 3.5 Mini Instruct | ~2.2 GB | 8 GB | Keeps full multi-head attention, so heavy on cache |
| Gemma 3n E4B | ~3.9 GB | 8 GB | Largest that fits an 8 GB iPhone |
Sizes are for the 4-bit quantised weights. "Needs" is the device memory the model wants, not the download.
Which one to start with
Apple Intelligence if you want an answer now. No download, no storage, and it cannot be beaten on time-to-first-use.
Qwen 3 4B if your phone has 6 GB or more. It is the best all-round model that still fits comfortably.
Llama 3.2 1B if your phone is older, or you want replies as fast as possible and the questions are simple.
What your phone can hold
Rough working ceilings:
| Device | Practical ceiling |
|---|---|
| iPhone 11–14 (4–6 GB) | 1B–1.7B with short context |
| iPhone 15 / 14 Pro (6–8 GB) | 3B–4B at 4K context |
| iPhone 15 Pro / 16 (8 GB) | 4B comfortable, 8B-class borderline |
Why a long chat costs more than the download
A model's weights are a fixed cost. The KV cache is not: it grows with the length of the conversation, and it is what actually ends a long chat.
A 3B model at 8K context can exceed 2 GB for cache alone — more than the weights. Two things keep that in hand, both on by default:
- Every entry defaults to a 4K context window.
- The cache is quantised to 8-bit, where the quality cost is negligible and the memory freed is what keeps a long conversation from being killed.
If a long conversation starts failing where a short one succeeds, the cache is the reason, not the model.
Speed
Expect the first generation after loading a model to be slower, because Metal compiles and caches its shaders on first use. Sustained generation also throttles thermally — a long reply on a warm phone is slower than the same reply on a cool one.
Running something bigger
No phone will hold a 30B model. A Mac will, and pairing one puts its models in your phone's model picker without the phone's size limits applying.