Google has released its latest multi-modal AI model called Gemma 4, designed to handle a variety of tasks involving reasoning, tool use, vision, and audio. What makes Gemma 4 stand out is its availability in different sizes, from small models suitable for personal hardware to larger ones for servers. Even at the higher end, the model remains responsive and practical for everyday use, thanks to recent architectural innovations claimed by Google.

Variety of Gemma 4 Model Sizes

Gemma 4 comes in four main versions, each tailored for different hardware capabilities. The smallest, E2B, has 2.3 billion effective parameters and a max context window of 128,000 tokens. Slightly larger, E4B, offers 4.5 billion parameters with the same context size. The 31-billion-parameter version is considered a dense model, with a larger context window of 256,000 tokens, but it’s quite hefty at around 62GB—probably not meant for personal machines.

There’s also a “mixture of experts” model called A4B, which activates 4 billion parameters out of a total of 26 billion, also with a 256,000 token context. Thanks to Gemma 4’s open licensing, the community has created many editions of these models, including compressed versions that use fewer bits for quantization, making them more manageable for smaller setups.

Testing Gemma 4 on Real Hardware

To see how well Gemma 4 performs, testing was done using a standard setup: an AMD Ryzen 5 3600 CPU, 32GB of RAM, and an Nvidia RTX 5060 GPU with 8GB VRAM. The models were run with a variety of prompts, including those testing vision capabilities, web search functions, and coding tasks.

For vision tasks, one prompt asked for a caption for an attached image, limited to three sentences and with an option to avoid editorializing. The models responded quickly and accurately, demonstrating their ability to interpret visual data. Other prompts aimed at web search functions requested detailed explanations about copyright laws and opinions of famous authors on classic movies. The models handled these well, providing clear and comprehensive answers.

A third type of prompt focused on coding, asking the models to help create a Python utility for customizing icons in executable files. This tested the models’ problem-solving skills and their ability to understand complex technical instructions. Overall, the models showed good responsiveness and versatility across all tasks, even on hardware that many users might have at home.

Thanks to architectural improvements, Gemma 4 manages to deliver strong performance without requiring the high-end hardware typically needed by large AI models. Its open licensing also encourages community development, making it easier for users to find versions tailored to their needs. Whether for personal projects or small business use, Gemma 4 offers a promising option for AI integration on a wide range of devices.

Inspired by

• https://www.infoworld.com/article/4156597/googles-gemma-4-shines-on-local-systems-both-big-and-small.html

Sources

• blog.google
• maartengrootendorst.com
• huggingface.co
• lmstudio.ai
• huggingface.co