xAI begins rolling out Grok-3 beta
xAI began rolling out Grok-3 beta with Think and DeepSearch; the initial release was limited to Grok products, with API access still pending.
What happened
Event details
On February 17, 2025, xAI began rolling out the Grok-3 family during a livestream. Its technical post dated February 19 described the models as an early preview that was still training. xAI said Grok-3 was trained on its Colossus cluster with roughly ten times the compute of the previous generation. The standard model had a one-million-token context window, and xAI reported scores of 73.2% on MMMU for image understanding and 74.5% on EgoSchema for video understanding. The launch also introduced Grok-3 (Think), Grok-3 mini (Think), and DeepSearch, which combined web retrieval and code tools with model reasoning. xAI said an early build code-named `chocolate` reached 1,402 Elo and led Chatbot Arena; these were launch-time vendor claims. TechCrunch confirmed that rollout had begun but noted that some models and features remained in beta, while the enterprise API was still weeks away.
Assessment
Why it matters
Grok-3 brought test-time reasoning, long context, and a retrieval agent into xAI's main product. Vendor benchmarks and a launch-time Chatbot Arena snapshot placed it among the leading models of the period, but training was ongoing, architecture and parameter count were undisclosed, and API access did not arrive with the product launch.
Availability
Access notes
The initial rollout prioritized X Premium+ users, with usage limits for other users. Grok-3 was not available through the API at launch and became generally available there on April 3, 2025.