Gemini 2.5 Pro Preview (Mar' 25) is an experimental iteration of Google's flagship multimodal model, first released on March 25, 2025. This version introduces a thinking capability, allowing the model to perform internal reasoning steps before providing a final response. This design is intended to improve performance on complex logic, multi-step problem-solving, and advanced coding tasks compared to previous non-thinking iterations.
The model maintains a 1 million token context window, enabling it to process and reason over massive datasets, including hours of video, extensive code repositories, and long-form documents. It is natively multimodal, supporting inputs across text, image, audio, and video formats. This allows for cross-modal reasoning, such as analyzing a video file alongside a specific code block to troubleshoot software behavior.
Upon its preview release, the model achieved high rankings on competitive benchmarks, including the LMArena leaderboard and STEM-focused evaluations like AIME 2025. It also features significant improvements in agentic capabilities, such as more reliable function calling and tool use, which facilitate its integration into autonomous workflows and complex developer environments.