Join daily and weekly newsletters to obtain the latest updates and exclusive content to cover the leading artificial intelligence in the industry. Learn more
Google More than his goal of “AI Global Assistant” is close to understanding context, planning and taking measures.
Today in Google I/O.TECH GIANT has announced its improvement in Gemini 2.5 Flash – it is now better in almost every dimensions, including standards for thinking, symbol and long context – and 2.5 Pro, including experimental enhanced thinking mode “, which allows professionals to consider multiple hypotheses before responding.
“This is our final goal of the Gemini application: Amnesty International, which is a personality, pre -emptive and strong,” said Dimis Hasabis, CEO of Google DeepMind, said in pre -shipment formulation.
“Deep Thinking” is impressive in the higher standards
Google has announced Gemini 2.5 Pro-what it considers the most intelligent models so far, with a window of millions-fill- In MarchAnd release it “I/o” coding version Earlier this month (with Hassan You call it “The best coding model we have ever built!”.
“We are really affected by the people created by people, from converting graphics into interactive applications to simulating the entire cities,” said account.
He pointed out that, based on Google’s experience with alphaThe artificial intelligence model responses improve when granted More time to think. This led to Deepmind scientists developed Deep Think, which uses the latest advanced research in Google to think and think, including parallel techniques.
Deep Think showed impressive degrees on the maximum standards of mathematics and coding, including Olympics of the United States of America for the year 2025 (Osamo). It also leads LiveCOOOOOBENCHA difficult standard for coding at the level of competition, and records 84.0 % on MmmuWhich tests a multimedia understanding and thinking.
“We take some extra time to make more border safety assessments and get more inputs from safety experts,” Hassabis added. (Meaning: And at the present time, the trusted laboratories are available via an interface programming interface for feedback before the ability is widely available.)
In general, the new PRO POPULAR CODING Leadorboard leads Webdev ArenaWith the ELO-degree that measures the level of relative skill for players in players games such as chess-from 1420 (intermediate to efficiency). It also leads across all categories LMARNA Topics, which evaluates artificial intelligence based on human preference.
Since its launch, “We have really been affected by what (users) created, from converting graphics into interactive applications to simulating the entire cities,” said Hassabis.
Important updates for Gemini 2.5 Pro, Flash
Today, Google has announced an improved 2.5 flash, a model for the work designed for speed, efficiency and low cost. 2.5 Flash has been improved in all areas in criteria for thinking, symbol, symbol and long context – Hassabis indicated that it is “only second” to 2.5 Pro on the LMARNA panel. The model is also more efficient, using 20 to 30 % of the distinctive symbols.
Google makes final adjustments to 2.5 flash based on the developer’s notes; It is now available for inspection in Google Ai Studio and Vertex AI and in the Gemini app. It will be generally available for production in early June.
Google brings additional capabilities to Gueini 2.5 Pro and 2.5 Flash, including outputing the original sound to create more natural conversation experiences, stipulating plans to support many speakers, thinking summaries and thinking budgets.
With the insertion of the original sound (in the inspection), users can direct the Gemini tone, its pulse and its style of speaking (Think: Directing the model to be microrami or Modelin when telling the story). Love Mariner ProjectThe model is also equipped with tools, allowing it to search on behalf of users.
Other early experimental sound features include an emotional dialogue, which gives the model the ability to discover feelings in the user’s voice and respond appropriately; The pre -emptive sound that allows it to adjust the background conversations; Thinking about the live application programming interface to support the most complex tasks.
The multiple new features in both Pro and Flash supports more than 24 languages, and models can quickly turn from one tone to another. Koraay Kavukcuoglu, CTO, written by Google DeepMind, and TULSEE DOSHI, First Director for Product Management at Google DeepMind, at Google Deep Mend, Tolsi, Senior Director of Product Management at Google DeepMind, in A. The blog was published today.
Moreover, 2.5 Pro and Flash now include thinking summaries in API Gemini and Vertex Ai. These “take the raw ideas of the model and organize it clearly with heads, main details and information about typical actions, such as when used tools”, explains Kavukcuoglu and Doshi. The goal is to provide a more organized and simplifying format for the model thinking and giving users interactions with Gemini is simpler to understand and correct.
Like 2.5 Flash, Pro is now equipped with “thinking budgets”, which gives developers the ability to control the number of symbols that the model uses to think before responding, or if they prefer, the potential of his thinking completely. This ability will be generally available in the coming weeks.
Finally, Google added the original SDK support to the McP -context protocol definitions in the Gemini Application interface so that the models can merge them more easily with open source tools.
Hassan said: “We live in a wonderful moment in history as it makes artificial intelligence a wonderful new future. It has been uncomfortable.”
https://venturebeat.com/wp-content/uploads/2025/05/Google.jpeg?w=1024?w=1200&strip=all
Source link