Science & Tech · AI · 6 hrs ago
ModelBest says LES runs a 35B AI model on iPhone at 20.51 tokens per second
ModelBest says its LES method runs a 35B mixture-of-experts model on a phone at more than 20 tokens per second.
A mixture-of-experts model uses different expert components across its layers.
LES fetches the next layer's experts from the phone's flash storage ahead of time, according to ModelBest.
The company reports a speed of 20.51 tokens per second on an iPhone, which it says is 6.07 times its baseline.
This matters because it shows how a large AI model can run on a phone, rather than relying only on remote computing.
ModelBest has open-sourced the code, making the method available for others to examine and use.
ModelBest says its LES method runs a 35-billion-parameter mixture-of-experts AI model on an iPhone at 20.51 tokens per second.
The reported speed is 6.07 times the baseline.
LES prefetches experts for the next model layer from the phone’s flash storage.
ModelBest has open-sourced the code.
- Who
- ModelBest.
- What
- It says its LES method runs a 35B mixture-of-experts model on an iPhone at 20.51 tokens per second.
- When
- The report was published on November 10, 2026.
- Where
- On an iPhone; the specific device location is not stated.
- Why
- LES prefetches next-layer experts from the phone’s flash storage. A reason for developing the method is not stated.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Model
- 35B mixture-of-experts AI model
- Method
- LES
- Reported speed
- 20.51 tokens per second
- Baseline comparison
- 6.07x
- Device
- iPhone
- Code
- Open-sourced