Gemma 3: Google DeepMind's Multimodal Open Model Family

The Gemma 3 Technical Report introduces Google DeepMind's Gemma 3, a new generation of lightweight open-source language models. These models offer enhanced capabilities including multimodal understanding, longer context windows (up to 128K tokens), and improved multilingual abilities. The report details architectural improvements focused on memory efficiency and training methodologies involving knowledge distillation and novel post-training recipes. It includes evaluations against other language models and the Gemini family, highlighting superior performance in mathematics, chat, and instruction following. The report also addresses safety, security, and responsible deployment, along with the model's carbon footprint. It includes analysis of memorization rates and safety policies.

keepSave to notecopy_all

docsAdd noteaudio_magic_eraserAudio OverviewschoolBriefing doc