rewrite this title DeepMind Debuts SIMA 2, Smarter AGI Powered by Gemini

This new model combines Gemini’s language and reasoning abilities with the skills needed to operate within virtual environments.

The first SIMA system relied on long recordings of people playing a wide selection of 3D games. Through this training, the model learned how to manage various tasks across many different game worlds, including some it had never practiced with.

SIMA 2 uses the Gemini 2.5 flash-lite model. DeepMind uses the term AGI to describe systems that can handle many types of thinking tasks, learn new abilities, and apply knowledge across different situations.

DeepMind’s researchers focus on “embodied agents”, which means the system has a virtual body. It can observe its environment and take actions in it, much as a person or robot would.

During a press briefing, DeepMind senior research scientist Joe Marino shared how much the model has improved. Joe Marino, senior research scientist at DeepMind, said:

SIMA 2 is a step change and improvement in capabilities over SIMA 1.

Another DeepMind researcher, Jane Wang, noted that SIMA 2 is not only about playing games. She explained that the agent must understand what is happening around it, make sense of the user’s request, and respond with basic reasoning.

Google recently brought its Gemini AI assistant to Maps in India, along with new tools focused on road safety and route updates. What did Miriam Daniel, Head of Google Maps, say? Read the full story.