Google Updates Gemini 3 Deep Think
Google's Gemini 3 Deep Think model has been updated with new multimodal reasoning capabilities. The AI can now convert sketches into 3D models, perform advanced visual and mathematical reasoning, and is being used to accelerate drug design.
- The drug design capabilities are an extension of Google DeepMind's AlphaFold 3, which can predict the structure and interactions of proteins, DNA, RNA, and ligands. Isomorphic Labs, an Alphabet company, has exclusive commercial access to AlphaFold 3 and is developing its own drug design engine that reportedly doubles AlphaFold 3's accuracy for some predictions. - The model's mathematical reasoning has achieved a gold-medal level performance at the 2025 International Math Olympiad. This is part of a broader "AI for Math Initiative" by Google to collaborate with research institutions to accelerate mathematical discovery. - Gemini 3 Deep Think's advanced reasoning is being positioned to compete with other specialized models like OpenAI's o1 and Anthropic's Claude. The model set a new record on the "Humanity's Last Exam" benchmark with a score of 48.4% without external tools and scored 84.6% on the ARC-AGI-2 reasoning benchmark. - The update is currently available to Google AI Ultra subscribers through the Gemini app. Google is also providing access to select researchers and enterprises via an early access program for the Gemini API. - This update is part of a broader strategy to integrate advanced AI into various Google products, including a universal AI assistant codenamed Project Astra. Project Astra aims to understand context and interact in real-time through multimodal inputs, with prototypes being tested on glasses and Pixel phones. - In the medical field, Google is researching a specialized AI called Articulate Medical Intelligence Explorer (AMIE). AMIE is designed to conduct diagnostic conversations and can now process visual information like ECG scans and dermatology images to improve its reasoning.
