Google AI Weekly Recap: Gemini 3.5, NotebookLM, and Gemma 4

Real-time Translation and Research Upgrades

Google introduced Gemini 3.5 Live Translate, an audio model capable of real-time speech translation across 70+ languages. This model is accessible via Google AI Studio. Additionally, NotebookLM received a significant update by integrating the Gemini 3.5 model and the Antigravity infrastructure harness. Users can now leverage 100+ curated software skills to perform research, generate structured reports, and create documents in formats like PDF and Microsoft Excel directly from natural language queries.

Model Architecture and Development Tools

Google released Gemma 4 12B, which utilizes a multimodal encoder-free architecture to improve processing speed. The company also introduced QAT (Quantization Aware Training) versions of the Gemma family, allowing for higher-quality results when models are compressed for local hardware. Furthermore, Diffusion Gemma was released as an experimental model that applies diffusion techniques to text generation, predicting blocks of text to optimize token usage. Developers should note that the Gemini CLI is being deprecated in favor of the Antigravity CLI, with a migration deadline of June 18th.

Data Science and Consumer Applications

Google Colab now supports CLI access, enabling developers to integrate Colab infrastructure into agentic AI workflows. The platform also added a visualization agent that allows users to generate dashboards directly within notebooks and share full-screen outputs without requiring viewers to navigate the entire notebook. Finally, Google Labs launched Dream Beans, a consumer application that connects to Gmail to generate personalized, cartoon-style narratives based on a user's email content. The company also announced Gemini SQL 2, a text-to-SQL model built on Gemini 3.1 Pro, aimed at improving natural language database querying.