Second-generation edge AI accelerator that runs vision and generative-AI workloads — including LLMs and vision-language models up to about 2B parameters — entir