Google releases a new generation of inference models.
On March 25, local time, Google announced the launch of the Gemini 2.5 series. The experimental version of Gemini 2.5 Pro is the first reasoning model launched in the series. Its highlights are as follows:
1. Powerful reasoning ability: The "thinking chain" mechanism has been built to improve the response quality through multiple rounds of logical deduction before generating answers, significantly enhancing the accuracy of handling complex problems. The deep analysis capabilities include information integration, logical argument construction, contextual detail grasp, and decision optimization.
2. Excellent benchmark performance: It ranked first with a 39-point advantage on Chatbot Arena, and also won the only championship in the three major fields of creative writing, instruction following, and long query; it scored the highest score of 18.8% in Humanity’s Last Exam, and also performed well in mathematics and science benchmarks such as GPQA and AIME 2025.
领英推è
3. Multimodal input support: As a native multimodal large model, it can handle multimodal inputs such as text, audio, images, videos, and large data sets, and can also understand the entire code repository of coding projects.
4. Extremely long context window: With an ultra-long context window of 1 million tokens, Google said it will soon expand to 2 million tokens, which can parse the entire "Lord of the Rings" series text, making it easier to handle long texts and complex tasks.
5. Efficient algorithm architecture: By optimizing the algorithm architecture, the response speed is increased by 40%, energy consumption is reduced by 25%, and the completion rate of complex logical tasks is increased by 65% compared with the previous generation, showing higher accuracy in vertical fields such as medical diagnosis assistance and legal document generation.
6. Excellent programming ability: The demonstration video shows that Gemini 2.5 Pro can create interactive charts based on prompt words, visualize complex data, or develop small games that are both designable and playable. It scored 68.6% in the aider polyglot code editing evaluation tool, surpassing models such as OpenAI o3-mini and Claude 3.7 Sonnet.