Michael Fukuyama

Google and NVIDIA have optimized the Gemma 4 family of open models across four variants (E2B, E4B, 26B, 31B) for efficient local deployment on NVIDIA hardware ranging from edge devices like Jetson Orin Nano to RTX PCs and the DGX Spark personal AI supercomputer. The models support reasoning, coding, agentic workflows, and multimodal capabilities including vision, video, and audio, with native function calling for tool use. NVIDIA has partnered with Ollama, llama.cpp, and Unsloth to provide accessible deployment paths, positioning these models as alternatives to cloud-based inference for developers and enterprises seeking on-device AI with local context access.