Google’s latest cook in AI models are now fresh out of the oven, and they are the Gemma 4 series, built from the ground up to be compact, fast, and capable enough to run efficiently on-device across a wide range of systems. And of course, it is made with close collaboration with NVIDIA to run the best with Team Green GPUs.
Consisting of 4 different model sizes for various tiers of hardware, as the smaller E2B and E4B models are tuned for ultra-efficient, low-latency performance at the edge with near-zero delay, while the larger 26B and 31B models are more geared toward heavier workloads like reasoning and developer-centric tasks, making them a strong fit for agentic AI setups.
All of them are capable of handling complex reasoning, supporting coding and debugging workflows, and come with built-in agent capabilities like structured tool use. On top of that, they expand into multimodal use cases with support for vision, video, and audio, while also allowing interleaved inputs – meaning you can mix text and images within the same prompt. They’re also multilingual out of the box, supporting over 35 languages and trained on more than 140.
With NeMoClaw and now Gemma, building local AI agents that can tap into personal files, apps, and workflows to automate tasks without relying on the cloud is now easier than ever.
On the deployment side, NVIDIA is smoothing things out by working with platforms like Ollama and llama.cpp, making it easier to run these models locally or even fine-tune them using tools like Unsloth Studio right from day one.
