Presenting Google Gemini 2.0, its updated AI model for the age of agents.
Google is launching the experimental Gemini 2.0 Flash model today, the first in the Gemini 2.0 family of devices. It is a workhorse model, offering improved performance and low latency at the forefront of its technology, at scale.
By presenting prototypes made possible by Google Gemini 2.0‘s built-in multimodal capabilities, it is also sharing the cutting edge of its agentic research.
Gemini 2.0 Flash
Gemini 2.0 Flash offers improved performance at remarkably quick response times, building on the popularity of Google most popular model for developers to date, 1.5 Flash. Notably, at double the speed, 2.0 Flash even beats 1.5 Pro on important metrics. New features are also included in 2.0 Flash. Flash 2.0 now supports multimodal output, such as natively created graphics with text and steerable text-to-speech (TTS) multilingual audio, in addition to multimodal inputs, such as photos, video, and audio. Additionally, it has native support for third-party user-defined functions, code execution, and tools like Google Search.
Google’s AI helper, Gemini 2.0, is accessible through the Gemini app
A chat-optimized version of 2.0 Flash experimental is also accessible to Gemini users worldwide as of right now by choosing it from the model drop-down on desktop and mobile web. It will soon be accessible through the Gemini mobile app. Users can have an even more useful Gemini helper with this upgraded model.
Google wants to extend Google Gemini 2.0 to additional Google products early in the upcoming year.
Google Gemini 2.0: Unlocking Agentic Experiences
Google Gemini 2.0 A new class of agentic experiences is made possible by Flash’s native user interface action capabilities, as well as additional enhancements like multimodal reasoning, long context understanding, complex improved latency, native tool usage, compositional function-calling, and instruction following and planning.
There are a lot of fascinating research opportunities in the field of practical AI agent application. With a number of prototypes that can assist individuals in completing activities and getting things done, Google is investigating this new frontier. These include Jules, an AI-powered code agent that can assist developers; the new Project Mariner, which investigates the future of human-agent interaction, beginning with your browser; and an update to Project Astra, Google’s research prototype investigating the potential of a universal AI assistant.
Although it is still in the early phases of development, Google is eager to observe how reliable testers utilize these new features and what insights it can draw from them so that it can expand their availability in future releases.
Project Astra: real-world agents employing multimodal comprehension
Trusted testers have been using Project Astra on Android phones to teach us since it first revealed at I/O. Their insightful comments have improved its understanding of the potential practical applications of a universal AI helper, including the ethical and safety ramifications. The most recent version, created with Google Gemini 2.0, has the following improvements:
- Improved conversation: Project Astra can now communicate in a variety of languages and mixed languages, and it can comprehend accents and unusual words better.
- Use of a new tool: Project Astra can now access Google Lens, Maps, and Search with Google Gemini 2.0, which makes it a more practical helper in daily life.
- Improved memory: Project Astra may now recall information better while maintaining your control. It is now more individualized for you because it can recall more of your previous talks and has up to 10 minutes of in-session memory.
- Improved latency: The agent can comprehend language at roughly the latency of a human conversation because to its enhanced streaming features and native audio comprehension.
Project Astra Glasses
It is trying to add these kinds of features to various form factors, including glasses, as well as Google products, like the Gemini app, its AI assistant. A limited group will shortly start testing Project Astra on prototype glasses as part of its efforts to broaden its trusted tester program.
Agents from Project Mariner can assist you in completing challenging jobs
Using Google Gemini 2.0, Project Mariner is an early research prototype that investigates how humans and agents will interact in the future, beginning with your browser. As a study prototype, it can comprehend and analyze data on your browser screen, including pixels and web elements like text, code, photos, and forms. It then uses this data to perform tasks for you using an experimental Chrome extension.
Project Mariner earned a state-of-the-art result of 83.5% when tested as a single agent setup against the WebVoyager benchmark, which evaluates agent performance on end-to-end real world online jobs.
Even though it’s not always precise and takes a while to do activities now, Project Mariner demonstrates that it is technically becoming feasible to move within a browser. This will quickly improve in the future.
Google is actively researching new risk categories and mitigations while keeping people informed in order to build this in a safe and responsible manner. Project Mariner, for instance, only allows you to type, scroll, or click in the tab that is now open on your browser. It also requests final confirmation before allowing you to do sensitive actions, such as making a purchase.
Jules: developers’ agents
Next, Google is investigating how Jules, an experimental AI-powered code agent that seamlessly integrates into a GitHub workflow, can help engineers. Under the guidance and oversight of a developer, it can address a problem, create a plan, and carry it out. It is working towards its long-term objective of creating AI agents that are useful in various fields, including coding.
Agents in games and other fields
Games have long been used by Google DeepMind to improve AI models’ ability to plan, obey rules, and reason. For instance, Google unveiled Genie 2, its AI model that can generate an infinite number of playable 3D environments from a single image, just last week. It has continued this legacy by leveraging Google Gemini 2.0 to create agents who can guide you through the virtual world of video games. In real-time conversation, it can provide advice for what to do next and make decisions about the game based just on the action on the screen.
It is working with top game creators like Supercell to investigate how these agents function, evaluating their capacity to decipher rules and challenges in a variety of games, from agricultural simulators like “Hay Day” to strategic games like “Clash of Clans.”
In addition to serving as virtual gaming partners, these agents can even use Google Search to link you to the abundance of online gaming resources.
Google is experimenting with agents that can assist in the real world by utilizing Google Gemini 2.0‘s spatial reasoning capabilities in robotics, in addition to investigating agentic capabilities in the virtual world.
Constructing responsibly in the age of agents
Google Gemini 2.0 At the vanguard of AI research, Flash and its research prototypes enable us to test and refine new features that will ultimately improve the usefulness of Google products.
It acknowledge the responsibility that comes with developing these new technologies, as well as the numerous safety and security concerns raised by AI agents. It is therefore using a methodical and exploratory approach to development, researching several prototypes, implementing safety training iteratively, collaborating with external experts and trustworthy testers, and carrying out thorough risk assessments as well as safety and assurance evaluations.
AI agents, Gemini 2.0, and beyond
With today’s releases, the Google Gemini model enters a new phase. It has achieved an important milestone in the Gemini era with the introduction of Gemini 2.0 Flash and the series of research prototypes investigating agentic potential. And as it moves closer to AGI, it is excited to keep safely investigating all the new options that are available.
