Explore the design, training methods, and unique challenges of creating fast, privacy-focused agentic AI models optimized for edge devices like smartphones.
In this presentation, we see the benefits and challenges of edge foundation models. We talk about model architecture to see how it can be leveraged to improve inference speed. We see post-training techniques and the new MOPD-centric recipe used by Liquid AI with Antidoom, Multi-Domain On-Policy Distillation (MOPD) and agentic Reinforcement Learning. Finally, we talk about agentic harnesses and how to co-design it with a model.