Musk Says xAI 2T Model to Finish Initial Training Next Week, Aiming to Surpass Kimi

On July 18, Elon Musk revealed on X that xAI's next-generation 2T (trillion) parameter model is expected to complete its initial training next week. He explicitly stated that the new model will be better in all aspects compared to the current 1.5T version, aiming to surpass the newly released Kimi K3, while maintaining inference speed and token efficiency close to the 1.5T model. This news quickly spread across the AI community, signaling a new escalation in the large model parameter race.

Key Details

Musk provided several clear expectations: regarding scale, it jumps from 1.5T to 2 trillion parameters; regarding performance, it is "better in all aspects" and specifically aims to beat Kimi K3 (according to Chuangyebang, Kimi K3 has 2.8 trillion parameters and has topped the Arena coding leaderboard); regarding efficiency, he promised that inference speed and token efficiency will be close to the 1.5T version (i.e., Grok 4.5), rather than sacrificing efficiency for scale. It should be noted that "2T" is the term used in Musk's original post, and the community generally refers to it as Grok 4.6 (though some reposts mistakenly called it Grok 4.5).

Community Reactions

Regarding the current 1.5T version and the new model, the community offered various subjective evaluations. @Elegant_Influence believes Grok 4.5 is highly efficient but not yet on par with Anthropic's Opus 4.8, placing hope in the 2T model (referred to by them as Grok 4.6) for a qualitative leap. @Arindam_1729 mentioned in testing that, subjectively, the current Grok is much better than Fable. Overall, the industry is waiting for the actual benchmark scores and deployment performance of this 2T model once its initial training is complete.

2026-07-18 ~ 2026-07-20 · 8 related posts

1 near-duplicate retellings: DanWahlin