Executive Summary
NVIDIA and Google have collaborated to optimize Google's new Gemma 4 family of open models for high-performance, local execution. These small, fast, and multimodal models are designed to run efficiently on NVIDIA hardware, from RTX PCs and DGX Spark personal supercomputers to Jetson edge devices. The partnership aims to accelerate the development of on-device agentic AI that can leverage local data for real-time reasoning, coding, and multimodal tasks.
Key Takeaways
* New Model Family: The announcement covers Google's Gemma 4, a new family of open models with E2B, E4B, 26B, and 31B parameter variants.
* Hardware Optimization: The models are specifically optimized to run on a wide range of NVIDIA hardware, including GeForce RTX GPUs, DGX Spark, and Jetson Orin Nano modules.
* Core Capabilities: Gemma 4 features strong performance in reasoning, coding, and native tool use (function calling) for agentic workflows.
* Multimodal & Multilingual: The models support interleaved multimodal input (mixing text and images in a single prompt) and have vision, video, and audio capabilities. They also support over 35 languages out-of-the-box.
* Targeted Use Cases: The smaller E2B and E4B models are designed for low-latency edge AI, while the 26B and 31B versions are suited for high-performance developer and agentic workflows on RTX PCs.
* Immediate Availability: Developers can access and run the optimized Gemma 4 models immediately using popular tools like Ollama, llama.cpp, and Unsloth Studio.
Strategic Importance
This collaboration reinforces NVIDIA's strategy to dominate the on-device AI market by ensuring the latest, most capable small models run best on its hardware, driving adoption of its RTX ecosystem for local AI agents beyond the data center.