Google has just rolled out its latest open-source AI models, the Gemma 4, designed for local use and boasting impressive specs. The four variants cater to different needs: two large models are optimized for high-performance tasks on powerful hardware, while smaller versions target mobile devices.

The large Gemma 26B Mixture of Experts and 31B Dense models can run unquantized on a single Nvidia H100 GPU, but quantization allows them to fit on consumer-grade GPUs. Meanwhile, the E2B and E4B effective models are fine-tuned for mobile devices with low memory footprints.

Google claims these new models will outperform their predecessors in terms of capabilities and efficiency. However, while they offer more control over AI usage, they require significant investment in hardware and expertise to fully leverage.

The move towards open-source licensing also signals a shift away from Google’s proprietary systems but may be seen as a step back for developers who prefer seamless cloud-based solutions. The company hopes that by providing these tools locally, it can foster innovation while maintaining some control over usage.