Gemini 2.5 Pro gets experimental enhanced inference mode 'Deep Think'
At Google I/O 2025, the annual developer conference held from May 21, 2025, Google announced new features for 'Gemini 2.5 Pro' and enhancements to 'Gemini 2.5 Flash.'
Google I/O 2025: Updates to Gemini 2.5 from Google DeepMind
In a post on May 21, 2025, the official Google DeepMind X (formerly Twitter) account announced, 'Introducing Deep Think in 2.5 Pro, a new enhanced reasoning mode that uses research into parallel thinking techniques to consider multiple hypotheses before responding. This allows you to more effectively handle incredibly complex math and coding problems.'
Deep Think in 2.5 Pro has landed. 🤯
— Google DeepMind (@GoogleDeepMind) May 20, 2025
It's a new enhanced reasoning mode using our research in parallel thinking techniques - meaning it explores multiple hypotheses before responding.
This enables it to handle incredibly complex math and coding problems more effectively. pic.twitter.com/xlyRq2Qu3R
According to Google, the Gemini 2.5 Pro Deep Think achieved a score of 49.4% in the 2025 USAMO math benchmark, which is one of the most difficult currently available. It also achieved 80.4% in the LiveCodeBench coding benchmark and 84% in the MMMU multimodal reasoning test, showing the highest performance in all indicators compared to the regular Gemini 2.5 Pro model.
'2.5 Pro DeepThink is truly on the cutting edge of technology, so we're taking more time to conduct more cutting-edge safety evaluations and get more input from safety experts,' said Tulsi Doshi, senior director of product management at Google. 'As part of that, we'll be providing it to trusted testers through the Gemini API for feedback before it's generally available.'
At the same time, Google announced that it has enhanced the performance of Gemini 2.5 Flash, a cost-effective model in the Gemini 2.5 series.
Performance comparisons show that the preview version of Gemini 2.5 Flash outperforms the previous model in most benchmarks, including inference and coding, and is particularly cost-effective. According to the announcement, the token usage of Gemini 2.5 Flash has been reduced by 20 to 30% due to this enhancement.
The enhanced Gemini 2.5 Flash is available in preview in Google AI Studio for developers, Vertex AI for enterprises, and the Gemini app for all users, with general availability expected in early June 2025.
in Software, Posted by log1l_ks