On July 17, Moonshot AI released the Kimi K3 large model, featuring a parameter scale of 2.8 trillion, a 1 million token context window, native visual understanding capabilities, and its proprietary KDA attention algorithm. Elon Musk immediately commented, “Impressive,” but the next day announced that xAI was training a new 2-trillion-parameter model aimed at surpassing Kimi. In response to this challenge, Moonshot AI humorously replied, “Welcome to the 2-trillion+ club,” highlighting its first-mover advantage with 2.8 trillion parameters. This back-and-forth exchange illustrates the technological rivalry between the two AI companies and reflects the rising influence of Chinese AI in the global large model competition.Article author and source: Wall Street Journal
On the early morning of July 17, Moonshot AI officially launched its next-generation large model, Kimi K3. With a parameter scale of 2.8 trillion and a 1 million token context window, the model natively supports visual understanding, and Moonshot AI describes it as "the most powerful model to date."
Multiple overseas AI evaluation rankings show that Kimi K3 ranks highly in tests such as code generation, long-text logical reasoning, and multi-document processing. The company has disclosed that Kimi K3 is powered by its proprietary KDA attention algorithm, featuring 896 independent expert modules and activating 16 units per computation, resulting in a 2.5x improvement in computational efficiency over the previous generation. The company plans to fully open-source the Kimi K3 model weights on July 27, granting commercial usage rights to developers worldwide.
On the day the reports about Kimi K3 were published, Musk commented on the article, describing the model as "impressive."
This is the second time this year that Musk has publicly affirmed the technical capabilities of the Kimi team. In March, Moonshot AI released the paper "Attention Residuals," introducing an "attention residual" mechanism designed to replace traditional residual connections in Transformer architectures to address the problem of information dilution in deep networks.
At the time, Musk shared the paper on X and commented, “Kimi’s work is impressive.” In response, Moonshot AI’s Kimi officially replied with a humorous remark: “Your rocket isn’t bad either!” This cross-platform exchange was seen as a mutual appreciation between top technologists and opened the door to subsequent positive interactions between the two sides.
However, shortly after praising Kimi K3, Musk began issuing strong statements remotely, showcasing the competitiveness of his own products.
On July 18, Musk posted on the social platform X, revealing xAI’s development plan for the Grok 4.6 large model, currently trained with 2 trillion parameters, aiming to complete its first full training cycle within a week. The new version is expected to outperform Grok 4.5, which has 1.5 trillion parameters, and Musk mentioned that this model has the potential to surpass Kimi in performance.
In 2023, Musk founded the AI company xAI, whose team members primarily came from OpenAI, Google DeepMind, Meta, and other institutions. The development of the Grok series of large models was launched simultaneously, with deep integration into the X social platform, Tesla, and SpaceX ecosystems.
In August 2023, xAI completed training of its first-generation foundational model, Grok-0, with 33 billion parameters. On November 5 of the same year, the first consumer-facing Grok was officially released, integrated into the X Platform Premium+ paid subscription system, and enabled real-time access to X’s live information feed, emphasizing a conversational style with minimal constraints.
In March 2024, xAI launched Grok-1, a model with 314 billion parameters, open-sourcing its full weights under the Apache 2.0 license; the same month, it introduced Grok-1.5, expanding the context window to 128K tokens and enhancing long-text reasoning capabilities. In April, Grok-1.5V was released, adding image and visual processing capabilities. In August of the same year, the multimodal Grok 2 was launched, alongside the lightweight Grok 2 mini, covering use cases such as general conversation, code generation, and image-text analysis. Simultaneously, xAI deployed its self-built Colossus supercomputing cluster for model training.
In February 2025, xAI released Grok 3, trained using the computational power of 100,000 NVIDIA H100 chips, with official disclosures indicating significant improvements in inference and multimodal performance. In July, Grok 4 was launched in two versions: the standard edition and Grok 4 Heavy, a multi-agent variant supporting up to a 256K context window, along with new capabilities in multi-voice voice interaction and basic video understanding. The originally planned Grok 3.5 version was subsequently canceled, with its development roadmap integrated into the Grok 4 series iteration.
In 2026, Grok entered the development phase of trillion-parameter models. In May, xAI completed training of the 1.5-trillion-parameter V9-Medium base model, incorporating industry data from the code tool Cursor to optimize for engineering and programming scenarios. On June 28, Grok 4.5, built on this base model, began internal closed testing at Tesla and SpaceX. On July 8, API access was fully opened to developers and paying users.
In response to Musk’s indirect challenge, the official Weibo account of Moonshot AI, “Moonshot Kimi,” directly tagged @ElonMusk and posted: “Welcome to the $2 trillion+ club.” This brief reply used the term “club” to signal an open and competitive stance, while implicitly highlighting that Kimi K3 already boasts 2.8 trillion parameters, clearly emphasizing its first-mover advantage. Some netizens remarked that when Musk said he might surpass Kimi, Kimi had already won.
Behind this "cross-distance challenge" lies a shift in the global AI industry's competitive landscape. Initially, the benchmark for large model technology was long dominated by Silicon Valley companies such as OpenAI and Anthropic. Over the past two years, Chinese large models like DeepSeek, GLM, and Kimi have continuously updated and iterated, rising to the center stage of global technological competition.
Public information shows that the legal entity behind Moonshot AI is Beijing Moonshot AI Technology Co., Ltd., which was registered in Beijing in 2023. The legal representative and CEO is Yang Zhilin. Yang Zhilin holds a bachelor’s degree from Tsinghua University and a Ph.D. in Computer Science from Carnegie Mellon University. He is a core inventor of the seminal pre-training frameworks Transformer-XL and XLNet, and previously worked at Google Brain and Meta AI.
Since its establishment, the company has successfully completed multiple large-scale funding rounds, with investors including Alibaba, Tencent, Meituan Dragonfly, Sequoia China, and Banyan Capital.
In October 2023, Moonshot AI launched its first consumer-facing intelligent assistant, Kimi Chat, introducing its initial 8K context base model and simultaneously enabling the capability to process ultra-long texts of up to 200,000 Chinese characters. In January 2025, the company released the K1.5 base model, enhancing performance in mathematical computation, code generation, and multimodal reasoning. In April of the same year, it open-sourced the 16-billion-parameter multimodal model Kimi-VL. In July, it unveiled the K2 series of base models with a trillion-parameter mixture-of-experts architecture, featuring a 256K-token context window and making them available as open-source. In January 2026, Moonshot AI released the K2.5 multimodal open-source model, integrating vision, code, and agent capabilities, alongside launching the programming tool Kimi Code. On July 17, it officially launched its current flagship base model, Kimi K3.
On July 17, Moonshot AI released the Kimi K3 large model, featuring a parameter scale of 2.8 trillion, a 1 million token context window, native visual understanding capabilities, and its proprietary KDA attention algorithm. Elon Musk immediately commented, “Impressive,” but the next day announced that xAI was training a new 2-trillion-parameter model aimed at surpassing Kimi. In response to this challenge, Moonshot AI humorously replied, “Welcome to the 2-trillion+ club,” highlighting its first-mover advantage with 2.8 trillion parameters. This back-and-forth exchange illustrates the technological rivalry between the two AI companies and reflects the rising influence of Chinese AI in the global large model competition.Source: Jiemian News