Elon Musk’s xAI Corp. late Monday night announced the launch of Grok-3, the latest in the company’s family of large language models.

The company says the AI model is a significant leap in power over its previous Grok-2 LLM that includes “reasoning models” to mimic human thinking.

During a livestream announcement, xAI researchers said that Grok-3 was trained using 10 to 15 times more compute power than Grok-2. The company launched a massive supercomputer training system named Colossus in September, built with 100,000 Nvidia Corp. H100 graphics processing units. It’s designed to bring new iterations of Grok online.

“Grok-3 across the board is in a league of its own,” Chief Executive Elon Musk said during the announcement, claiming that the model can outperform models from OpenAI and China’s DeepSeek based on early testing in math, science and coding.

Grok-3 includes two main reasoning models: Grok-3 Reasoning beta, a large complex model, and Grok-3 mini Reasoning, a small fast model capable of generating quick answers. With the models enabled on xAI’s chatbot, the models will reveal their “thinking” as they do step-by-step reasoning during complex science, mathematics and programming questions.

This release comes at a time when other companies have started to release reasoning models that break down complex tasks into smaller tasks and then attempt to fact-check themselves before offering solutions. The objective is to provide a better result. Example models include those developed by competing companies such as OpenAI’s o1 and o3-mini reasoning, DeepSeek’s R1 and Google LLC’s Gemini 2.0 Flash Thinking Experimental.

“We should emphasize that this is kind of a beta, meaning that you should expect some imperfections at first, but we will improve it rapidly, almost every day,” Musk added.

To access the reasoning capabilities of the Grok-3 models, users can turn on “Think” to have it reason through their queries. And for more difficult questions, they can activate “Big Brain” mode, which xAI says is best suited for more complicated queries that involve reasoning in math, science or coding.

The reasoning mode for Grok-3 can also be paired with a search capability called “DeepSearch,” which takes longer but causes the model to scan the internet for relevant knowledge and incorporate it into its answers. XAI said using DeepSearch will result in more relevant, detailed responses.

The addition of this deep internet research capability will mean that xAI’s models will join rival models which have similar features including OpenAI and Google. Perplexity AI Inc., the creator of an AI search engine, bakes deep internet research directly into its service when presenting search answers.

Grok-3 will also receive a voice mode that will allow it to respond to queries verbally. Although the feature will not be available at launch, Musk said it will be rolled out in a week or so. The voice mode for Grok will be more than voice-to-text, Musk added. It will understand tone, inflection and pacing, and “it will be like talking to a person.”

Image: xAI