Next Generation Reasoning Model
Gemini 3 Pro is the next generation in the Gemini series, designed for complex problem-solving. This natively multimodal model is capable of comprehending vast datasets across text, audio, images, video, and code repositories.
It features an optional 'Deep Think' mode to enhance complex decision-making during inference. The model can process up to 1 million tokens in its context window and supports outputs of up to 64,000 tokens.
Architecture and Infrastructure
Gemini 3 Pro utilizes a sparse mixture-of-experts (MoE) transformer architecture, routing tokens dynamically to subsets of parameters to optimize capacity and computation. It was trained extensively on Google's Tensor Processing Units (TPUs) using reinforcement learning techniques to leverage multi-step reasoning and theorem-proving data.
Core Capabilities & Safety
- Reasoning & Logic: Outperforms previous iterations across agentic performance, algorithmic development, and complex problem-solving.
- Multimodal Integration: Processes diverse data sources, including documents, code repositories, audio, and visual inputs, within a 1M token context.
- Enhanced Safety: Thoroughly evaluated for content safety, showing measurable improvements in unjustified-refusals and tone over Gemini 2.5 Pro.
- Frontier Safety Checked: Tested under Google DeepMind's Frontier Safety Framework, identifying no critical capability levels reached in cyber security or misalignment.
Sources & Credentials
The information and model specifications regarding Gemini 3 Pro are sourced directly from the official Google DeepMind Model Card.