TradingKey - On July 18, Eastern Time, SpaceX ( SPCX) and Tesla ( TSLA) CEO Elon Musk revealed on the social platform X that the training of Grok 4.6, the new-generation large model developed by his AI company xAI, has entered its final stage and initial training is expected to be completed next week. (Note: Musk's original words on July 18 were "next week," which counts as this week based on the July 20 post.)
[Source: X]
According to Musk, this new model with 2 trillion parameters is "superior in every way" to the current 1.5 trillion-parameter Grok 4.5 version, while its inference speed and token efficiency will remain basically on par with the previous generation. In a reply to a user, he stated that this model has the "potential to surpass" Moonshot AI's newly released Kimi K3.
This statement comes shortly after Chinese AI startup Moonshot AI released Kimi K3. On July 17, Moonshot AI officially launched its next-generation large model, Kimi K3, with a parameter size of 2.8 trillion, making it currently the world's largest open-source model by parameter size. On the day of its release, Musk commented "impressive" in the comments section of a related review report.
Since its release, Kimi K3 has performed impressively on third-party evaluation platforms. In the Arena AI frontend coding leaderboard, K3 topped the global rankings with a score of 1,679, surpassing Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. Meanwhile, Moonshot AI officially admitted that K3's overall user experience still has a significant gap compared to these two top-tier closed-source models.
The official Kimi account responded to Musk's statement, posting on the social platform Weibo: "@ElonMusk Welcome to the '2 Trillion+' club." It is reported that the full model weights of Kimi K3 will be open to global developers before July 27.
[Source: Weibo]
Currently, xAI has not yet announced the official launch time, full training plan, or technical evaluation metrics for Grok 4.6. Looking back at the R&D process, xAI completed the training of its 1.5-trillion-parameter base model V9-Medium in May 2026, which is regarded as the core foundation of Grok 4.5.
In terms of computing power support, the Grok 4 series models are trained using a hyper-scale AI computing cluster codenamed "Colossus." In January 2026, xAI announced that Colossus 2 had become operational, equipped at launch with approximately 555,000 GB200 and GB300 series GPUs, equivalent to the computing power of approximately 1.4 million H100 GPUs. xAI previously disclosed that the ultimate goal for Colossus 2 is a capacity of 2 gigawatts (GW), with plans to add more than 400 megawatts (MW) of computing power in June 2026, incorporating over 220,000 new GB300 chips, with an ultimate deployment target of 1 million GPUs.
Industry analysis indicates that while the total number of parameters in a large language model is an important reference for measuring model scale, the model's actual comprehensive capability also depends on multiple factors such as training data quality, algorithmic framework, inference architecture, computing power supply, and post-training optimization. Whether xAI can maintain the operational efficiency of the previous generation while expanding the model scale will be key to whether Grok 4.6 can pose a substantive challenge to Kimi K3.
Find out more
Comments