A Saarland University preprint tests personal adaptation of the SmolLM3 language model on an iPhone 17 Pro. Short training runs on users’ movie-tagging examples achieved similar reported accuracy to a server experiment. Longer runs ran into battery exhaustion and reduced throughput as the phone heated up. The work uses small trainable adapters rather than retraining all model weights, and its findings come from one phone and software setup.
A single preprint from Saarland University tests adapting the SmolLM3 language model on an iPhone 17 Pro. The model has 3 billion parameters, its learned numerical settings. The researchers use LoRA, a method that trains small adapter components while keeping base weights unchanged, and low-precision weights. Training runs through MLX Swift, software for machine-learning computation. The work is a preprint rather than a peer-reviewed publication.[1]
The personalisation experiment uses movie-tagging data from 100 users. The median user has 36 examples, with training taking approximately 3 minutes. Across 3,302 queries, the researchers report accuracy of 48.6 per cent on the phone and 48.8 per cent on the server, against 36.6 per cent before adaptation. These are results from the researchers’ experiment on one device and software setup.[1]
A longer run with 405 examples over three passes took 1.74 hours and used 47 per cent of the battery. With 550 examples, training took 3.07 hours and used 80 per cent. A larger run stopped when the battery ran out. Heat reduced throughput by approximately half. Adding pauses lowered hourly training-step completion. The phone’s iOS operating system also required the application to remain in the foreground with the screen on for the graphics processor’s computation to continue.[1]
