Musk xAI Launches Grok-2, Mini Version With Improved Reasoning Skills

Elon Musk's xAI startup has unveiled Grok-2, the latest version of its language model, boasting significant improvements in reasoning abilities and academic comprehension while outperforming rival AI chatbots in benchmark tests.

Released in an early preview, Grok-2 offers “significant improvements” on the previous Grok-1.5, offering more intuitive and versatile outputs across a variety of tasks

The model’s reasoning abilities have been improved, with Grok-2 able to correctly identify missing information in a piece of content and reason through a sequence of events.

An early version of the model under the codename "sus-column-r" outperformed both Anthropic’s Claude 3.5 Sonnet and OpenAI’s GPT-4 on the industry standard LMSYS Chatbot Arena.

Credit: xAI

The newly launched model is currently available in beta on X, locked behind the $8 a month X (Twitter) Premium subscription.

Accessible through a new look interface, Grok-2 is designed to work as an AI assistant solution on X in a similar vein to ChatGPT, offering text and visual understandings. The model can be used to answer user queries, collaborate on creative writing or generate code snippets.

Grok-2 will be made available to enterprise users through the xAI API “later this month.”

Enterprise users can feel safer using the model thanks to newly introduced security features, including multi-factor authentication, advanced billing analytics and improved team and user management preventing unwanted access to the model.

Related:Musk AI Chatbot Under Fire for Sharing False Election Information

XAI says the release of Grok-2 has positioned the year-old startup “at the forefront of AI development.”

Musk, who founded xAI to compete with OpenAI, was full of praise for the team behind the model, saying their rate of progress was “excellent.”

He teased Grok-2’s launch last month and suggested that its successor, Grok-3, would drop around the end of the year and would be on par with the highly coveted but as yet released GPT-5 from OpenAI.

Grok-2 Mini: Small But Mighty

In addition to Grok-2, xAI launched Grok-2 mini, a smaller version of the model that boasts impressive performance despite its size.

XAI says the smaller model “offers a balance between speed and answer quality.”

Despite its size, the mini Grok-2 outperforms the full-sized Grok 1.5 model — and even outperformed models far larger, including Claude 3 Opus and Google’s Gemini 1.5 Pro on benchmark tests including MMLU, HumanEval and MMLU-Pro.

Credit: xAI

Grok-2 mini will also be available on X, allowing users to choose which model they want to use.

It’s also coming to the enterprise API at the end of August, with the smaller model designed to be cheaper to run while still offering capable enough performance.

Related:Musk AI Chatbot Under Fire for Sharing False Election Information

What’s Next for XAI?

In addition to building what it hopes will be the world’s largest supercomputer to train Grok-3, xAI is turning its attention to different multimodalities.

The startup said it’s going to release a preview of Grok’s multimodal understanding on the X app “soon.”

This would mean Grok-2 models would be able to handle inputs beyond text, including potential images, videos and maybe even audio.

Details on the exact modalities weren’t disclosed, but Musk wants future Grok models to be some of the most powerful AI models ever produced, which would require it to handle inputs beyond solely text.

Musk said he believes the next Grok model needs as many as 100,000 H100 GPUs to train effectively.

He’s working with Nvidia, Dell and Supermicro to build a massive supercomputer in Memphis, Tennessee.

However, some residents are opposed to the plan, expressing concern the supercomputer’s massive power needs could threaten residential electricity supply and overburden infrastructure.