Elon Musk’s xAI Holdings Corp. has debuted a new large language model, Grok 4, that’s optimized for reasoning tasks such as generating code.

The LLM’s late Wednesday launch followed a turbulent week for the company. A few hours before the product announcement, X Chief Executive Officer Linda Yaccarino resigned. Earlier, the artificial intelligence chatbot built into the social network generated a series of antisemitic replies to user posts.

X, formerly Twitter, became part of xAI through a $33 billion all-stock acquisition in March. The social network’s chatbot is powered by the Grok LLM series.

Grok 4 was trained on Colossus, a supercomputer that xAI launched in Memphis last year. The system featured 100,000 graphics cards when it came online in September. According to xAI, that number topped 200,000 in May and will eventually reach 1 million.

The latest model is designed to process prompts with up to 256,000 tokens worth of text and images. It can analyze charts, generate code, solve math problems and perform related tasks. According to xAI, the LLM features significantly better reasoning capabilities than the previous-generation Grok 3 model.

The company evaluated Grok 4 using Humanity’s Last Exam, an AI benchmarking dataset known for its complexity. It contains 2,500 questions spanning multiple scientific fields. Grok 4 solved more than 44% of the questions without using any external applications such as search engines. OpenAI’s Deep Research tool, an AI agent that uses its o3 reasoning model, achieved a 26.6% score.

“Grok 4 is at the point where it essentially never gets math/physics exam questions wrong, unless they are skillfully adversarial,” Musk wrote in a post on X. “It can identify errors or ambiguities in questions, then fix the error in the question or answer each variant of an ambiguous question.”

It’s available to developers through xAI’s application programming interface. The company will also offer an enhanced version, Grok 4 Heavy, that it says uses multiple AI agents to process queries. The agents each answer the user’s question, compare their responses and output the best one.

Grok 4 Heavy will be available through a subscription called SuperGrok Heavy that is priced at $300 per month. The plan will also provide early access to other new xAI products. 

In an X live stream, Musk said the company plans to launch a version of Grok 4 optimized for programming tasks next month. In September, xAI will introduce a second edition with expanded multimodal capabilities. A third version of Grok 4 built to generate videos will launch a few weeks after that.

In the longer term, xAI plans to connect the LLM to scientific applications of the kind that Tesla Corp. uses to design vehicles. That includes computational fluid dynamics software, which helps engineers simulate a car’s aerodynamics. Additionally, xAI hopes to make Grok 4 available on major public cloud platforms. 

Image: Unsplash