Members of the Artificial Intelligence Division (AI) of X, XAI, and the owner of the platform, Elon Musk, have announced their new family of language models Grok 3which will rely on the Deep Search search engine and promises to overcome OpenAi GPT-4O in the Aime and GPQA tests for doctoral reasoning.
The presentation of the new model of X has taken place this Monday, in a retransmission that has lasted about an hour and in which the most intelligent of the planet “, in the words of the owner of the owner of the Platform, Elon Musk.
This Improved Grok version 2 Includes the characteristics of its predecessor, with improvements in sections such as chat or reasoning, as well as the generation of images; And adds others, such as the ability to reflect on the mistakes he makes, in order to achieve logical coherence.
The X artificial intelligence division has also indicated that it planned to launch Grok 3 in 2024. However, a few more months have been taken to profile it, try new capabilities and make it a “much more capable reasoning model than Grok 2”, according to Musk said during the presentation.
This coincides with what he advanced a few days ago, when he said in X that he had completed the preventive phase of the model “with 10 times more processing capacity than Grok 2”. However, in the meeting they have suggested that “may” have a capacity 15 times greater.
First, the owner of X has indicated that Grok and, more specifically, the family of Grok 3 models, which will hide their reasoning processes, is composed of Grok-3 reasoning y Grok-3 mini reasoning, that responds more quickly, although sacrifices the precision of your answers for it, as it has suggested.
He also pointed out that this family has been created “with the mission of understanding the universe” and that, therefore, it is still “in a kind of beta”, since some of its abilities are still “irregular”, such as the Voice mode “Literally, in 24 hours you will see improvements,” other members of the XAI team have added during the retransmission.
“We believe that having the best training model is not enough. The best should think as a human being. You have to contemplate all possible answers, self -assess and verify mistakes,” added those responsible for Grok 3, which have indicated that they have indicated that they have indicated that You can solve tasks related to STEM disciplines.
In this sense, XAI has also affirmed that Grok 3 exceeds GPT-4O in reference tests such as Use Math Olympiad (AIME), which evaluates the performance of a model in a sample of mathematics questions, and GPQA, that is, it evaluates models using physics, biology and chemistry problems Doctoral level.
De Ese Modo, Grok 3 Reasoning Y Grok 3 Mini Reasoning They can “think” carefully In the problems, similar to the models of reasoning such as O3-mini of OpenAI and R1 of Deepseek. Also, the first exceeds the best version of o3-mini-highat several reference points, such as Aime 2025.
Also, the new XAI reasoning models support a new function in the application for iOS and Android called DeepSearch. Like Google and Openai’s proposals, which have a similar name, this search tool collects information from the Internet and the X application to offer an exhaustive summary according to the consultation made.
Grok 3 will first reach the subscribers of the X LEVEL+ Xalthough other functions will be restricted to a new plan that XAI has called Supergrok. This unlocks additional reasoning consultations, DeepSearch and offers unlimited generation of images.
Finally, those responsible for Grok have indicated that in a few weeks the Grok 3 models will be available through the XAI Application Programming Interface (API), together with Deepsearch, and that also also They plan to release the Grok 2 source code in the coming months.
https://t.co/hEfQ31gANQ
— xAI (@xai) February 18, 2025