Elon Musk announced on July 18 that xAI’s upcoming large language model, anticipated to be named Grok 4.6, will conclude its initial training run next week. This announcement follows the launch of Moonshot AI’s Kimi K3, a 2.8-trillion-parameter model, two days earlier on July 16.
Kimi K3 has been positioned as the largest open-weight AI model released to date and is scheduled for public release of its full model weights on July 27. Musk mentioned that Grok 4.6 aims to surpass Kimi K3 while maintaining performance metrics similar to those of Grok 4.5, which is based on a 1.5-trillion-parameter model known internally as V9.
Testing carried out by Artificial Analysis rated Kimi K3 at 57 on intelligence benchmarks, compared to a score of 54 for Grok 4.5. Grok 4.5’s per-task cost is approximately $0.31, significantly less than Kimi K3’s cost of $0.94, with Musk stating that the 2T model is intended to close this performance gap while keeping cost efficiency.
Independent evaluations indicated that Kimi K3 competes well against leading U.S. models and excels in web interface engineering, as reported by the BBC. In addition, Business Insider noted that Moonshot AI’s API pricing for Kimi K3 is lower than those of both OpenAI and Anthropic.
The development of the 2T model is part of xAI’s larger scaling strategy, which includes plans for 6-trillion and 10-trillion-parameter models referred to as Grok 5, currently under training on the Colossus 2 supercluster. As of April, seven models were reportedly in concurrent training on the Colossus 2 system.
xAI also continues to refine Grok Build, a coding agent introduced in May, which has reached version 0.2.101 as of July 13 and operates from the terminal while coordinating work with parallel subagents. This tool remains available in beta for subscribers of xAI’s SuperGrok Heavy service, priced at $300 per month.
