Google's Gemma 4 Runs Frontier AI On A Single GPU

Google DeepMind released Gemma 4, four open-weight models that fit on an 80GB H100 GPU and are Apache 2.0 licensed.

The family ranges from a 31B dense model to 26B MoE and 4B/2B edge models, supporting multimodal input, native function calling, and long context windows.

Nvidia and AMD published day-zero optimizations and Nvidia provided deployment tools like NIM and NeMo for enterprise and edge workflows.

Early community tests reported inference speed and fine-tuning compatibility issues for some configurations.

The Apache 2.0 license and broad hardware support make Gemma 4 easier to deploy on-premises for enterprises.