Sure, here you go; this data is using my Mac. I haven't been able to test it on an iPhone, as I don't have the 16, but I believe it should also work with the iPhone 16.
Context and Setup
Device used: Mac Studio M1 Ultra 64GB with macOS 15.0.1
Models tested: Mistral7B-Instruct V0.3:
- INT4
- FP16
- Decoding algorithm: Greedy
First Run (Model Adaptation):
This is the initial execution when the app is first launched, and Core ML adapts the model to your device’s hardware. This adaptation process is time-consuming as Core ML optimizes the model to leverage the device’s capabilities in future runs.
| Configuration | Execution Type | Time for First Token (s) | Generation Speed (tokens/s) |
|---|---|---|---|
| INT4 | First Run (Model Adaptation) | 32.81 | - |
| INT4 | Subsequent Runs - First Token | 2.62 | 4.55 |
| INT4 | Additional Iteration | 0.32 | 3.90 |
| FP16 | First Run (Model Adaptation) | 87.28 | - |
| FP16 | Subsequent Runs - First Token | 23.17 | 5.93 |
| FP16 | Additional Iteration | 7.52 | 4.09 |
