TomatoLM runs tomato-0.1 — an 8.99M-param GPT trained from scratch in Termux on a Mali GPU — inference 100% client-side in your browser
This model is not the smartest, and that wont be its goal any time soon. the goal is a locally run model for low end devices.