Google Just Released Gemini 2.0 Flash and Pro for Users and Developers
Google has released another round of major AI model announcements, updating its Gemini offerings across the board to provide users and developers with AI engines that the company says are more capable and robust. Thanks to the growth of DeepSeek and new OpenAI models, the pace of AI development is not slowing down.
First, the Gemini 2.0 Flash model, which launched in December for a select few, is now rolling out to everyone, so you’ll see it in Gemini apps on desktop and mobile (it actually started rolling out last week, so you may have already used it). Flash models are designed to be faster and lighter without sacrificing too much performance.
Google is also making the Gemini 2.0 Flash Thinking Experimental available for testing by all users. This is another “reasoning” model similar to what we saw in ChatGPT , where the AI displays its thinking as it goes along – with the intention of producing more accurate and more transparent results.
There is also a version of this model available to all users with access to the included applications: Google Search, Google Maps and YouTube. It will return real-time information from the Internet, as well as links to Google Maps data (including travel times and location information), as well as information pulled from YouTube videos.
Finally, regarding Flash models, Google is making Gemini 2.0 Flash-Lite available to developers. This is Gemini’s most cost-effective model and will appeal to those who build tools with Gemini while maintaining high levels of performance for processing a variety of multimodal inputs (text, images, etc.).
Professional models
Next up is an even more powerful experimental model, Gemini 2.0 Pro—slightly slower than its Flash counterparts, but better at thinking, writing, coding, and problem solving. This model is now in experimental form for developers and for all users who pay $20 per month for Gemini Advanced .
“It has the fastest coding performance and ability to handle complex queries, and better understanding and grounding of global knowledge, than any model we’ve released to date,” says Google . It can also accept two million tokens per invite, which averages out to about 1.4 million words—roughly the Bible, twice.
That’s double the performance of the 2.0 Flash models, and Google has provided some benchmarks as well. In the overall MMLU-Pro test, we got scores of 71.6 percent, 77.6 percent, and 79.1 percent, respectively, for the Gemini 2.0 Flash-Lite, 2.0 Flash, and 2.0 Pro, compared to 67.3 percent for the 1.5 Flash and 75.8 percent for the 1.5 Pro.
Similar improvements were seen in other AI tests, with Gemini 2.0 Pro Experimental scoring 91.8 percent on a leading math test. This compares to 90.9 percent for Flash 2.0, 86.8 percent for Flash-Lite, 86.5 percent for 1.5 Pro, and 77.9 percent for 1.5 Flash.
As is typical when running such artificial intelligence models, details about the training data used, the risks and inaccuracies of hallucinations , and energy requirements are sparse. While Google claims the new Flash models are the most effective, all of its latest models are better than ever at analyzing feedback and preventing potential security breaches.