Google has unveiled its latest AI advancement, the Gemma 4 models, offering significant enhancements over previous iterations. With a focus on local processing, these models come in four optimized sizes: 26B Mixture of Experts and 31B Dense, designed for high-performance GPUs like the $20,000 Nvidia H100. Google has responded to developer concerns by eliminating its custom licensing restrictions. The two larger models can run efficiently on consumer GPUs when quantized, enhancing accessibility. Additionally, Gemma 4 includes mobile-optimized variants, Effective 2B (E2B) and Effective 4B (E4B), achieving low memory usage for devices such as smartphones and Raspberry Pi. Both are engineered for “near-zero latency” and improved performance, surpassing the capabilities of Gemma 3. Google claims that Gemma 31B will rank third among top open AI models, demonstrating the robustness and cost-effectiveness of these new iterations, making them a prime choice for developers focused on AI innovation.
Source link
Share
Read more