Google releases Gemma 2 2B, ShieldGemma and Gemma Scope

AI Impact Summary

Google has released Gemma 2 2B, ShieldGemma, and Gemma Scope, expanding the Gemma model family. Gemma 2 2B is a 2.6B parameter version of Gemma 2, suitable for on-device use, while ShieldGemma provides safety classifiers to filter harmful content. Gemma Scope offers a suite of sparse autoencoders for Gemma 2 2B and 9B, enabling efficient inference. These releases leverage bfloat16 for inference and integrate seamlessly with the Hugging Face ecosystem via Transformers, demonstrating a focus on accessibility and optimization for diverse deployment scenarios.