Google has just introduced its Gemini 2.5 model family, which is the most powerful version of its AI assistant to date. The Gemini 2.5 models are capable of thinking and reasoning before providing a response, thereby enhancing performance and accuracy.
To develop them, Google leveraged the base model of its previous version, Gemini 2.0 Flash Thinking, and enhanced it with optimized training, thus reaching new levels of performance.
As the first preview of this new family of models, the technology company has launched Gemini 2.5 Pro Experimental, which they classify as their «most advanced model for complex tasks». It features a context window of 1 million tokens (soon to be 2 million) and can understand large datasets of all types (text, audio, images, video, or code) thanks to its multimodal capabilities.
This model is already available in Google AI Studio and in the model selector of the Gemini app for users of Gemini Advanced, and it has been announced that it will soon be activated in Vertex AI. Additionally, Google has announced that, in the coming weeks, they will disclose pricing to enable the use of this version with higher speed limits for large-scale production.
From now on, the technology giant will integrate thought capabilities directly into all its models so that «they can address more complex problems and support even more capable and context-sensitive agents».
The 2.5 family is based on the two key qualities of the Gemini models: their multimodal capability and extensive context windows. Additionally, it features advanced reasoning capability, a quality that has proliferated among cutting-edge models from other companies, such as ChatGPT’s deep research, Perplexity’s deep research, or DeepSeek R1.
As Google explains, «in the field of AI, a system’s ability to reason goes beyond simple classification and prediction. It refers to its ability to analyze information, extract logical conclusions, incorporate context and nuances, and make informed decisions».
The company has shared the results obtained by its Gemini 2.5 Pro model in various performance tests and a comparison with other relevant AI models.
Regarding its coding capabilities, Google claims to have made a significant leap in quality compared to the Gemini 2.0 family. «2.5 Pro excels in the creation of visually appealing web applications and agentic code applications, in addition to code transformation and editing. In SWE-Bench Verified, the industry standard for agentic code evaluations, Gemini 2.5 Pro scores 63.8% with a customized agent configuration».
The technology company published the following video showcasing how the reasoning capabilities of Gemini 2.5 Pro can be utilized to create a video game by generating executable code from a simple request.
Photo: Google
Your email address will not be published. Required fields are marked *
Δ