Introducing Gemini 2.0: our new AI model for the agentic era

December 11, 2024 | Source: Google DeepMind Blog

Tags: Gemini 2.0, Google DeepMind, agentic AI, multimodal, Gemini API, Project Mariner

Google launched Gemini 2.0 with Flash as the first model — 2x faster than Gemini 1.5 Flash, with native multimodal output, improved tool use, and agentic capabilities demonstrated through Project Mariner browser-use and Project Astra.

Details

Google DeepMind unveiled Gemini 2.0 in December 2024, marking a significant architectural advance over the 1.5 generation. Gemini 2.0 Flash, the initial released model, achieves roughly double the inference speed of Gemini 1.5 Flash while maintaining comparable quality across reasoning, coding, and multimodal benchmarks.\n\nThe most distinctive new capability is native multimodal output: Gemini 2.0 can generate images and audio directly, not just process them as inputs. This enables richer agentic interactions where the model responds with the appropriate modality rather than always generating text.\n\nGoogle demonstrated agentic applications through two research prototypes: Project Mariner, a browser-using agent that navigates web interfaces to complete multi-step tasks, and Project Astra, a real-time universal assistant with persistent memory. Both use Gemini 2.0 as their underlying model.\n\nDevelopers got immediate access via Gemini API and Google AI Studio at no cost (with rate limits). The 2.0 Pro and Ultra variants were announced for 2025. The release was Google's clearest competitive answer to GPT-4o and set the stage for a major agentic AI capabilities race in 2025.