NUPIAO Staff Writer
Elon Musk and Moon’s Dark Side are trading verbal jabs again, and it’s heating up.
On July 21, according to reports from 21st Century Business Herald, Huang Zhenxin, head of enterprise business at Moon’s Dark Side, fired back after Musk claimed that “Grok 4.6 will soon surpass Kimi.” “Maybe this performance (of K3) gave them a bit of a shock,” Huang said. “Now, we’re waiting to see. We hope they come out and really wrestle with us.”
Huang noted that while they don’t know Musk personally, the back-and-forth has been intense. He called Musk’s recognition “pretty precious” for the team.
This isn’t the first time Musk has noticed this Chinese company. Back in March, Moon’s Dark Side published a paper on “Attention Residuals,” proposing a new mechanism to replace traditional residual connections in Transformer architectures. Musk shared it on social media, commenting, “Kimi’s work is impressive.” The company playfully replied, “Your rockets are pretty cool too!”
Fast forward four months, and on July 17, Moon’s Dark Side quietly launched Kimi K3, a 2.8 trillion-parameter mixture-of-experts model—no flashy launch event, but it still set the industry buzzing. Musk quickly dropped a “Impressive” under a review post, again nodding to their tech chops.
At the same time, Musk spilled details on Grok 4.6. “Our 2 trillion parameter model beats our 1.5 trillion one across the board and will finish initial training next week. It might overtake Kimi, though inference speed and token efficiency are similar to our 1.5 trillion version (Grok 4.5).” Many see this as Musk throwing down the gauntlet.
Moon’s Dark Side shot back on Weibo: “Welcome to the ‘2 trillion+’ club.”
Why Musk zeroes in on Kimi? It’s all about the tech. Moon’s Dark Side says Kimi K3 uses a self-developed KDA hybrid linear attention mechanism and attention residual tech, with 896 independent expert modules, activating just 16 per inference—boosting compute efficiency by 2.5 times over its predecessor.
It natively supports a 100 million-token ultra-long context and multimodal visual understanding, shining in code generation, complex logical reasoning, batch document processing, and agent execution. For instance, on the Frontend Code Arena benchmark, Kimi K3 scored 1679, beating Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol.
Moon’s Dark Side plans to fully open-source Kimi K3’s weights on July 27, offering commercial licenses to global developers. API pricing: $3 per million input tokens and $15 per million output tokens—higher than before but still cheaper than closed-source flagships.
A new report from CITIC Securities highlights Kimi K3 as the world’s first nearly 3 trillion-parameter open-source large model, signaling that China’s models are closing the gap with overseas leaders. It’s not just about size; it’s about capability, inference efficiency, and real-world deployment. As domestic models evolve, demand for computing power could spark a new infrastructure boom, with opportunities in domestic model makers and the computing chain.
Interestingly, after Kimi K3’s launch, user demand skyrocketed beyond expectations. On July 20, the company announced a halt to new consumer subscriptions, diverting all compute power to existing users to ensure their experience stays smooth.
Moon’s Dark Side says it’s racing to scale up computing power. As new resources come online, it will gradually reopen subscriptions. For future users, the company plans to split Kimi’s core benefits (including Kimi Web, Kimi App, and Kimi Work) from Kimi Code, aiming to better match compute resources and keep user experience top-notch.
Back in June, Huang Zhenxin told NUPIAO and other media that Kimi’s B2B business share keeps growing, with sectors like internet, finance, manufacturing, education, and healthcare becoming key clients, while overseas business is booming too.
Huang stressed that the company’s core focus remains on deep investment in model architecture, including pushing the boundaries of Scaling Law.
Moon’s Dark Side sits firmly in China’s top tier of AI startups. Just two months after founding, it raised nearly 2 billion yuan in angel funding, with heavyweights like Sequoia China and ZhenFund on board. In February 2024, it secured over $1 billion, hitting a $2.5 billion valuation; by August, another $300 million pushed it to $3.3 billion. In December 2025, it closed a $500 million Series C, valuing it at $4.3 billion, with over 10 billion yuan in cash.
In the first two months of this year, the company closed three rounds—$500 million, $700 million, and $700 million—lifting its valuation from $10 billion to $18 billion, with Alibaba, Tencent, and 5Y Capital doubling down. Including a $2 billion round in May, Kimi has raised over 37.6 billion yuan cumulatively, making it the most-funded domestic AI startup. For context, MiniMax has raised about 15 billion yuan, and Zhipu around 13 billion. Market buzz says it’s in talks for a Pre-IPO round, eyeing a Hong Kong listing in as little as six months.
Despite the applause for Kimi K3, gaps remain with GPT and Claude in multimodal deep understanding, ultra-long complex agent loops, and bottom-level compute ecosystem adaptation. Huang admits the long-term goal is “to explore the ceiling of intelligence and really go toe-to-toe with those top three model companies overseas.”
