The company of Elon Musk, Xai, released her latest flag model he, Grok 3, late Monday, along with new skills in Grok apps for iOS and online.
Grok, Xai’s response to models like GPT-4o of Openai and Google Gemini, can analyze images and answer questions, and strengthens a number of features on Musk’s social network, X. Grok 3, who has been in Developing for several months was optimism planned for release in 2024, but lost that deadline.
Monday is an ambitious start.
Xai has used a large data center in MEMFIS – a data center containing about 200,000 GPU – to train Grok 3. In a post on X, Musk claimed Grok 3 was developed with “10x” more computing than Grok 2, its predecessor, and with a set of expanded training data seemingly includes records of court cases.
“Grok 3 is an order of size more capable than Grok 2,” Musk said during a broadcast live on Monday. “(Is a) He is maximally seeking the truth, even if this truth is sometimes contrary to what is politically accurate.”
Grok 3 is a family of models, to be correct – not just one. A smaller version of Grok 3, Grok 3 Mini, answers questions faster at the cost of some accuracy. Not all patterns and related features are still available (and some are in beta), but participation begins on Monday.
Xai claims that Grok 3 GPT-4o beats in standards, including AIME, which evaluates the performance of a model in a sampling of mathematics questions, and GPQA, which evaluates models using doctoral level physics, biology and chemistry problems . An early version of Grok 3 also scored competitive in the Chatbot Arena, a crowd test that pits different models of one against each other and has the users to vote for their favorite answers, according to Xai.
Two changes of Grok 3, Grok 3 Reasoning and Grok 3 Mini Reasoning, can “think carefully” problems, similar to “reasoning” as Openai’s O3-Mini and Chinese company AI Company Deepseek’s R1. Reasoning patterns thoroughly control themselves before yielding results, which helps them avoid some of the traps that normally travel models.
Xai claims that Grok 3 Reasoning exceeds the best version of O3-O3-O3-II-High in some popular standards, including a newest mathematics reference point called AIME 2025.
Reasoning patterns can be achieved through the Grok application. Users can ask Grok 3 to “think”, or – for more difficult questions – to use “Big Brain” mode for the reasoning it uses additional computing. Xai describes reasoning patterns as most suitable for questions related to mathematics, science and programming.
Musk said that, in the Grok app, some of the “thoughts” of reasoning patterns have been darkened to prevent distillation, a method used by model developers to draw knowledge from another model. Recently, Deepseek was charged with distillation of OpenAi models to create his own.
Groku’s reasoning models support a new feature in the Grok app called Deepsearch, Xai’s response to “deep research” tools with him as deep research of Openai. Deepsearch scans the Internet and X to analyze the information and give an abstract in response to a question.
The X -‘Mier+ Tier subscribers (22 dollars a month) will first receive Grok 3, and other features are connected after a new plan Xai is calling Supergrok. Ecedeated with $ 30 a month or $ 300 a year (if the leaks are trusted), Supergrok unlocks the additional reasoning and deep search questions, and throws in unlimited image generation.
In the future – as soon as about a week from now – the Grok app will gain a “voice fashion”, said Musk, who will give Grok models a synthesized voice. A few weeks after that, Grok 3 models will reach the API of the XAI enterprise, along with the ability of Deepsearch.
Xai plans to open the Grok 2 source in the coming months, Musk said.
“Our general approach is that we will open the source of the last version (of Grok) when the next version is completely out,” he continued. “When Grok 3 is ripe and durable, which is likely to be within a few months, then we will open with open source Grok 2.”
When Musk announced Grok for approximately two years ago, he raised he as clumsy, non-filtered and anti- “woke up”-in general, ready to answer controversial questions that other systems of he do not will. He handed over some of those promises. He said to be vulgar, for example, Grok and Grok 2 would fortunately force, promoting colorful languages that you would probably not hear from the chatgt.
But Grok models before Grok 3 was protected by political entities and would not cross certain boundaries. In fact, one study found that Grok bowed to the political left on topics such as transgender rights, diversity programs and inequality.
Musk has blamed the behavior in Grok’s training data – public websites – and pledged to “move Grok closer to neutral politics”. It is not yet clear if Xai achieved that goal – and what the consequences may be.
