Musk says xAI’s 2T-parameter model will finish training next week

Elon Musk disclosed on July 18 that xAI's next large language model — a 2-trillion-parameter system expected to be called Grok 4.6 — will complete its initial training run next week, a direct response to Moonshot AI's launch of Kimi K3 just two days earlier.kucoin+2

Musk said the new model outperforms Grok 4.5, which is built on a 1.5-trillion-parameter foundation model known internally as V9, across all metrics. He added that the 2T model may surpass Kimi K3 while maintaining generation speed and token efficiency close to that of Grok 4.5.kucoin+1

The Kimi K3 Challenge

Moonshot AI unveiled Kimi K3 on July 16 at 2.8 trillion parameters, making it the largest open-weight AI model ever released. The Beijing-based startup timed the launch ahead of the World Artificial Intelligence Conference in Shanghai, with full model weights scheduled for public release on July 27.tmtpost+4

According to testing by Artificial Analysis, Kimi K3 scored 57 on intelligence benchmarks compared to 54 for Grok 4.5, though Grok 4.5's per-task cost of roughly $0.31 is about one-third of Kimi K3's $0.94. Musk's stated goal for the 2T model is to close that performance gap while preserving the cost advantage.kucoin+1

The BBC reported that independent assessments from Artificial Analysis indicate Kimi K3 performs comparably to leading U.S. models, achieving top marks in web interface engineering. Business Insider reported that Moonshot AI's Alphabet Inc. ↘ 1.24% API pricing for K3 undercuts both OpenAI and Anthropic.bbc+1

xAI's Scaling Ambitions

The 2T model represents one step in xAI's broader scaling roadmap. Grok 4.5, the 1.5-trillion-parameter model currently in private beta at SpaceX and Tesla ↘ 1.30%, was announced by Musk on June 28 and is reportedly performing "close to Claude Opus, and in some cases surpasses it."gigazine+2

Beyond that, xAI has even larger systems in training on its Colossus 2 supercluster, including 6-trillion and 10-trillion-parameter variants designated as Grok 5. Seven models were reportedly training concurrently on the cluster as of April.mindstudio+1

Separately, xAI continues developing Grok Build, a coding agent launched in May that operates from the terminal and delegates work to parallel subagents. The tool reached version 0.2.101 on July 13 and remains in beta for subscribers of xAI's $300-per-month SuperGrok Heavy tier.ciodive+2