Just recently, Chinese AI company Moonshot AI officially unveiled its next-gen open-source large model, Kimi K3. Tesla CEO Elon Musk even chimed in under a related review, calling it “Impressive.” But then, on July 18th, Musk took to social media to announce that his own AI company is training a new model with 2 trillion parameters, and it “might surpass Kimi.”
Facing Musk’s virtual “challenge,” Moonshot AI’s official Weibo account, “Moonshot AI Kimi,” directly @ElonMusk with a playful post: “Welcome to the 2 Trillion+ Club.”
On the early morning of July 17th, Moonshot AI officially launched its next-gen large model, Kimi K3. With a whopping 2.8 trillion parameters, a 100 million token context window, and native visual understanding, the company calls it “the most capable model to date.”
Multiple overseas AI benchmark rankings show Kimi K3 leading in tests like code generation, long-text logical reasoning, and multi-document processing. Moonshot AI reveals that Kimi K3 runs on its self-developed KDA attention algorithm, boasts 896 independent expert modules, activates 16 units per calculation, and is 2.5 times more efficient in compute power usage than its predecessor. They plan to fully open-source Kimi K3’s model weights on July 27th, granting commercial access to developers worldwide.
The day Kimi K3 news broke, Musk dropped a comment under the report, calling it “Impressive.”
This is the second time this year Musk has publicly praised Kimi’s tech chops. Back in March, Moonshot AI published a paper on “Attention Residuals,” introducing a mechanism to replace traditional residual connections in Transformer architectures, tackling the issue of information dilution in deep networks.
At the time, Musk reposted the paper on X and commented, “Kimi’s work is impressive.” The official Kimi account then cheekily replied, “Your rockets aren’t bad either!” That friendly banter was seen as a mutual respect between top tech minds, and it set the stage for more good-natured exchanges.
But this time, right after praising Kimi K3, Musk started “talking trash” from a distance, showing off what his own product can do.
On July 18th, Musk posted on X, revealing xAI’s plan to train a 2 trillion parameter Grok 4.6 model, with the first full training run expected to wrap up in a week. This new version is an upgrade over the 1.5 trillion parameter Grok 4.5, and Musk hinted it could outperform Kimi.
Musk founded xAI in 2023, pulling talent from OpenAI, Google DeepMind, and Meta. The core product, the Grok series of large models, is tightly integrated with the X platform, Tesla, and SpaceX ecosystem.
In August 2023, xAI finished training its first base model, Grok-0, with 33 billion parameters. By November 5th of the same year, the first user-facing Grok launched, tied to X’s Premium+ subscription, with real-time access to X’s data feed and a no-holds-barred conversation style.
March 2024 saw xAI release Grok-1, with 314 billion parameters, fully open-sourced under Apache 2.0. That same month, Grok-1.5 rolled out, expanding the context window to 128K tokens. April brought Grok-1.5V, adding image processing. By August, the multimodal Grok 2 dropped, along with a lighter Grok 2 mini for general chat, coding, and image analysis. Meanwhile, xAI’s Colossus supercomputer cluster fired up for training.
In February 2025, xAI launched Grok 3, trained on 100,000 Nvidia H100 chips, boasting big leaps in reasoning and multimodal performance. July saw Grok 4, with standard and Grok 4 Heavy multi-agent versions, supporting up to 256K context windows, plus multi-voice speech and basic video analysis. The planned Grok 3.5 was scrapped, with its tech folded into the Grok 4 series.
2026 marked Grok’s entry into trillion-parameter territory. In May, xAI trained a 1.5 trillion parameter V9-Medium base model, optimizing for engineering and programming tasks. On June 28th, Grok 4.5, built on that base, started internal testing at Tesla and SpaceX. By July 8th, it opened API access to developers and paid users.
This time, Moonshot AI kept its response short: “Welcome to the 2 Trillion+ Club.” The word “club” signals an open, competitive stance, with the clear subtext that Kimi K3 already has 2.8 trillion parameters, highlighting its first-mover advantage. As one netizen put it, when Musk said “might surpass Kimi,” the real winner was Kimi itself.
This “virtual face-off” also reflects a shift in the global AI competitive landscape. For a long time, the benchmark for large models was set by Silicon Valley giants like OpenAI and Anthropic. But lately, Chinese models like DeepSeek, GLM, and Kimi have been iterating fast, taking center stage in global tech competition.
Public info shows Moonshot AI, also known globally as Moonshot AI, was registered in Beijing in 2023. Its legal rep and CEO is Yang Zhilin, who holds a bachelor’s from Tsinghua University and a Ph.D. in computer science from Carnegie Mellon. He’s a core inventor of Transformer-XL and XLNet, and previously worked at Google Brain and Meta AI.
Since its founding, the company has secured multiple rounds of big funding from investors like Alibaba, Tencent, Meituan Dragon Ball, Sequoia China, and Gaorong Capital.
In October 2023, Moonshot AI launched its first consumer AI assistant, Kimi Chat, with an initial 8K context model and a 200,000-character long-text feature. By January 2025, they released the K1.5 base model, upgrading math, code, and multimodal reasoning. In April of that year, they open-sourced the 16-billion-parameter Kimi-VL model. July saw the release and open-sourcing of the trillion-parameter K2 series with a 256K context window. In January 2026, they launched the K2.5 multimodal open-source model, integrating vision, code, and agent capabilities, along with the Kimi Code tool. And on July 17th, they unveiled their flagship Kimi K3.
