The world of artificial intelligence is evolving at an impressive speed, and Google has its say in this battle. The Mountain View firm has just announced the arrival of Gemini 2.5 Pro, a new AI model that takes the time to "think" before responding. Execution speed is not the priority of this new LLM, which presents itself as the most intelligent to date. A first version, still experimental, promises increased performance thanks to this "thinking" capacity. improved.
Google is touting the merits of Gemini 2.5 Pro based on the results in certain benchmarks, such as LMARena. The American giant's model ranks number 1 in this ranking based on the preferences of human users, with a significant lead over its pursuers. It particularly excels in complex tasks requiring advanced reasoning and obtains interesting scores in recognized benchmarks in mathematics, science and logic, such as GPQA or AIME 2025. Each time, Gemini 2.5 Pro outperforms o3 mini or OpenAI’s GPT-4.5, as well as Anthropic’s Claude 3.7 Sonnet and DeepSeek R1. The programming capabilities of this new model also take a leap forward. Gemini 2.5 Pro can generate code for complex web applications, assist developers in transforming existing code or, as Google shows, create the executable code for a video game. from a simple sentence.
How does Gemini 2.5 Pro stand out?
The big new feature of Gemini 2.5 is its ability to "Think" before generating a response. In AI jargon, the Californian firm points out that "reasoning" is not limited to classifying information or predicting an outcome. It involves the system analyzing data in depth, drawing logical conclusions, integrating context and nuances, and ultimately making more informed and relevant decisions. Google says it has combined a significantly improved base model with new training techniques to achieve this level. The goal is to integrate this reasoning capability into all future models.
Gemini 2.5 Pro is natively multimodal, meaning it can simultaneously understand and process text, images, audio, video, or lines of code. This legacy of previous versions is complemented by a "context window" of one million tokens, which will soon be increased to two million. Concretely, this allows it to analyze huge amounts of information such as (entire books, hours of video, large code bases) to solve complex problems.
Availability
The good news for the curious is that it is already possible to experiment with Gemini 2.5 Pro for free via Google AI Studio. This may be of particular interest to developers.
Subscribers to Gemini Advanced users can also experience Gemini 2.5 Pro now in the Gemini app, on desktop and mobile. This announcement comes just days after the launch of Gemma 3.
0 Comments