xAI, Elon Musk's company, released the Grok3 reasoning model, claiming it surpasses o1 and R1

Article source: Phoenix Technology

Image source: Generated by Wujie AI

On February 18, Musk's xAI released the large model Grok 3.

Previously, Musk claimed that Grok 3 is "the smartest artificial intelligence on Earth."

In the AIME'24 math ability test, Grok-3 scored 52 points, significantly higher than DeepSeek-V3's 39 points. In the GPQA scientific knowledge assessment, Grok-3 led with an excellent score of 75 points compared to DeepSeek-V3's 65 points. Additionally, in the LCB Oct-Feb programming ability test, Grok-3 surpassed DeepSeek-V3 with a score of 57 versus 36.

According to xAI, Grok 3 utilized 100,000 Nvidia H100 chips during training, a several-fold leap compared to the 15,000 GPUs used for its predecessor Grok 2.

The biggest highlight of Grok 3 is the introduction of a "Chain of Thought" (CoT) reasoning mechanism. This mechanism enables Grok 3 to solve complex problems step-by-step like humans, significantly enhancing the model's ability to handle complex queries and deliver more coherent and logical responses. Musk stated at the launch event that this reasoning mechanism allows Grok 3 to outperform ChatGPT-5 and Google's Gemini Ultra in logical reasoning tests, enabling it to generate novel ideas and solutions.

The Grok large model has been integrated into X, Musk's social media platform, allowing users to access the Grok chatbot on X.