Google officially announced Gemini 2.0 on Wednesday, the company’s latest upgrade to its flagship AI model. Specifically, Google is rolling out “Gemini 2.0 Flash experimental,” succeeding Gemini 1.5 Flash. Google’s Flash models are its “lightweight” models, designed for tasks that don’t require the most powerful AI models possible, and focus more on efficiency. Still, Google says Gemini 2.0 Flash not only improves upon Flash models like Gemini 1.5 Flash, but also more powerful models like Gemini 1.5 Pro.
How does Gemini 2.0 Flash compare to other models?
Table of Contents
Google says that 2.0 Flash beats both 1.5 Flash and 1.5 Pro in a number of categories, including General MMLU-Pro benchmarking, three different coding benchmarks, a factuality test, two math benchmarks, reasoning, two image benchmarks, and video benchmarks. Some of these wins were close to 1.5 Pro’s performance, however others showed significant improvement, such as a 7.5 point increase in the Natural2Code benchmark, or a nine-point increase in the HiddenMath benchmark. 1.5 Pro still beats 2.0 Flash in audio benchmarking (40.1% vs. 39.2%) and long context benchmarking (82.6% vs. 69.2%).
In addition to these improvements, Google says 2.0 Flash supports new multimodal outputs, such as AI-generated images combined with text and text-to-speech audio. Plus, it can pull in Google Search, run code, in addition to other third-party functions.
Where you’ll see Gemini 2.0 Flash
You’ll probably be seeing a lot of Gemini 2.0 Flash—whether you know it or not. The company announced that it will be using Gemini 2.0 for Search, specifically AI Overviews. The initial rollout of Google’s AI search summaries was, unequivocally, a hot mess. Nevertheless, the company is expressing optimism about the feature: Google says Gemini 2.0 will enable AI Overviews to handle more complicated topics and multi-step queries, as well as new functions like advanced math, multimodal questions (i.e. queries from text, images, documents, etc.) as well as coding.
2.0 Flash is also coming to the Gemini app. In fact, it’s already available on desktop and the mobile web experience. You’ll just need to choose the model from the drop down menu before testing it out.
Google wants AI to do the work for you
Google is advertising 2.0 Flash as part of its “agentic era.” What this means is that Google wants its products to do more on your behalf, whether that be analyze a question or your surroundings, or to actually complete a task for you. The company says they’re working on updates to Project Astra, Google’s research department responsible for developing a “universal AI assistant”; Project Mariner, a Chrome extension that utilizes AI to help you while browsing the web; and Jules, the company’s AI-powered agent to help developers write code.
Google also highlighted a new feature it calls “Deep Research,” an AI-powered research assistant that aims to analyze subjects and generate reports for you. You prompt the bot with a topic or question you want to investigate, and it develops a research plan for you to approve or revise. Once approved, it scrapes the internet for sources, and puts together a full report you can export to Google Docs. Like AI Overviews, it includes links to the sources it pulls from, so you can review them on your own.