October 2025 Summaries

Infineon developed a family of compact language models designed to run entirely on its PSOC™ Edge microcontrollers using less than 32 MB of memory, targeting privacy-sensitive, ultra-low-power applications such as wearables, IoT products, and embedded systems. Built over four to five months, the models range from 8 million to 25 million parameters and were optimized through staged sequence training, data-processing refinements, fine-tuning, and human-feedback methods to support on-device language understanding and assistant functions. The team used Lightning’s on-demand GPU and CPU resources, checkpointing, experiment tracking, multi-GPU scaling, and hardware telemetry to overcome limits in its internal compute capacity and to monitor utilization without additional instrumentation. Infineon estimates these capabilities reduced a development process that might otherwise have taken 1.5 to 2 years, demonstrating that functional language models can operate offline on highly constrained hardware while emphasizing energy efficiency and data privacy.