The company AI d’Elon Musk, XAI, late Monday, published its latest flagship model AI, Grok 3, and has unveiled new capacities for Grok iOS and Web applications.
Grok, XAI’s response to models like the GPT-4O of Openai and the Gemini of Google, can analyze the images and answer questions, and feed a certain number of features on the social network of Musk, X. Grok 3 , which has been in development for several months, was optimistic optimistic about released in 2024, but missed this deadline.
Monday is an ambitious launch.
Xai uses a huge data center in Memphis containing around 200,000 GPU to form Grok 3. In a job On X, Musk said that Grok 3 has been developed with “10x” (or more) more computing power than its predecessor, Grok 2, using an enlarged training set which includes deposits of trialS – and more.
“Grok 3 is a more capable order of magnitude than Grok 2,” said Musk during a live presentation on Monday. “”[It’s a] The search for truth to the maximum, even if this truth is sometimes in contradiction with what is politically correct. »»
Grok 3 is a family of models, to be precise. A smaller version of Grok 3, Grok 3 Mini, answers questions faster at the price of a precision. All the models and related features of Grok 3 are not yet available (some are in beta version), but they started to take place on Monday.
XAI claims that Grok 3 bats GPT-4O on references, in particular likes (which assesses the performance of a model on a sample of mathematical questions) and the GPQA (which assesses the models using problems of physics, biology and chemistry at the doctorate level). A first version of Grok 3 also marked competitively in Chatbot Arena, a crowdsource test which oppose different models of AI against each other and passes on their favorite answers, according to XAI.
Two models of the new Grok 3 family, the Grok 3 reasoning and the Grok 3 Mini reasoning, can carefully “reflect” on the problems, similar to the “reasoning” models like the O3-Mini and the Chinese IA company Deepseek’s R1’s R1 . The reasoning models try to withdraw from the facts before giving results, which helps them to avoid some of the traps which normally stumbles.
XAI claims that the Grok 3 reasoning exceeds the best version of O3-Mini-O3-Mini-High-on several popular benchmarks, including a more recent mathematics repository called AIM 2025.
These reasoning models are accessible via the GROK application. Users can ask Grok 3 to “think” or – for more difficult queries – take advantage of the “Big Brain” mode for reasoning that uses an additional calculation. XAI describes models of reasoning as best suited to mathematics, science and programming issues.
Musk said that some of the “thoughts” of reasoning models are hidden in the Grok application to prevent distillation, a method used by AI model developers to extract knowledge from other models. Recently, Deepseek has been accused of distilling OpenAi models to create his own.
Grok’s reasoning models underpin a new feature in the Grok application called Deepsearch, XAI’s response to research tools fueled by AI as in-depth research of Openai. Deepsearch Scanne Internet and X to analyze the information and deliver a summary in response to a question.
The premium + subscribers of X ($ 50 per month) will first have access to Grok 3, and other features will be integrated into a new plan than the call of XAI Supergrok. At a price of $ 30 per month or $ 300 per year (If we believe the leaks), Supergrok unlocks additional reasoning and Deepsearch requests, and launches an unlimited generation of images.
In the future – from a week for about a week – the Grok application will win a “vocal mode”, said Musk, which will give Grok models a synthesized voice. A few weeks after that, the Grok 3 models will be available via the XAI Enterprise API, as well as the Deepsearch capacity.
XAI plans to open the Open Grok 2 in the coming months, said Musk.
“Our general approach is that we will open the latest version of the latest version [of Grok] When the next version is completely extinguished, “he continued. “When Grok 3 is mature and stable, which is probably in a few months, we will open the source.”
When Musk announced Grok about two years ago, he presented the AI model as upset, not filtered and anti -“woken up” – in general, willing to answer controversial questions that other AI systems do will not do it. He held part of this promise. Said to be vulgar, for example, Grok and Grok 2 would readily force, spitting a colorful language that you would probably not hear a Chatppt.
But the Grok front models Grok 3 covered on political subjects and would not cross certain limits. Actually, a study noted that Grok was leaned over the political left on subjects such as transgender rights, diversity programs and inequalities.
Musk blamed the behavior of Grok training data – Public web pages – and prompted “Bringing Grok closer to politically neutral politics”. It is not yet clear if Xai has achieved this objective and what could be the consequences.
