NVIDIA Co-Optimizes Google Gemma 4 for Edge Agent Deployment: 31B Dense to 2.3B Edge Across RTX, DGX Spark, and Jetson

NVIDIA Developer Blog·medium signal

NVIDIA announced April 2 that it co-optimized Google's Gemma 4 family for local and edge agentic AI deployment. The four-model bundle — 31B dense, 26B MoE (128 experts), E4B, and E2B — supports reasoning, code generation, agent tool use, and multimodal input. The 31B model runs on DGX Spark's 128GB unified memory while E2B (2.3B effective parameters) targets Jetson Orin Nano for robotics. Models deploy via vLLM, Ollama, llama.cpp, and NeMo Automodel with commercial-friendly licensing.