Google DeepMind releases Gemini 3 Pro model card with Deep Think mode

Google DeepMind has published the model card for Gemini 3 Pro, its next-generation multimodal reasoning model released in November 2025. The document outlines the model's architecture, training data sources, and evaluation results compared to previous versions.

• Gemini 3 Pro is a sparse mixture-of-experts transformer with native support for text, audio, images, and video inputs.
• It features an optional "Deep Think" mode designed to enhance complex problem-solving performance at inference time.
• The model supports a context window of up to 1 million tokens and generates outputs of up to 64K tokens.
• Training utilized Google's Tensor Processing Units (TPUs) and JAX software, incorporating reinforcement learning with multi-step reasoning data.
• Evaluation results as of November 2025 indicate that Gemini 3 Pro significantly outperforms Gemini 2.5 Pro across benchmarks for reasoning, multimodal capabilities, agentic tool use, and long-context tasks.