Moonshot AI's new open model is catching up to Anthropic and OpenAI, raising questions about the future of the AI race
Kimi K3's performance shows that Chinese AI capabilities could be catching up to those of U.S. frontier labs.
A new Chinese AI model is causing a stir, having landed in the top spot of Arena's Frontend Code leaderboard and hinting that the capability gap between the U.S. and China could be closing faster than expected.
The model, called Kimi K3 and developed by Beijing-based startup Moonshot AI, is the world's largest open model with 2.8 trillion parameters, according to the AI lab. The model weights will be made available by July 27, Moonshot shared in a blog post. That means "anyone, anywhere, can download it and build on top of it for free," Mark Malek, CIO of Siebert Financial told MarketWatch. Early comparisons put Kimi K3 "right in the conversation" with top U.S. closed models, Malek said.
The release of Kimi K3 added further fuel to a selloff in tech stocks, notably chip names. The Nasdaq Composite COMP dropped 1.4% on Friday as the market digested the implications of a Chinese model achieving such results with less compute resources than the U.S.
There were already long-term anxieties regarding rising AI capital expenditures and circular financing methods among the hyperscalers, according to Steve Hou, head of research at Silicon Data. "If you have open models that are coming onto the scene, catching up in capability, then that's going to put pricing pressure on the frontier labs," Hou told MarketWatch.
Hou drew comparisons to last year's DeepSeek moment, which caused Nvidia and other chip stocks to tumble on fears that compute efficiencies would reduce demand for AI infrastructure. Over a year later, Big Tech companies have only increased their AI investments, and could be on track to spend over $1 trillion cumulatively in 2027, according to Morgan Stanley. Kimi K3 could potentially deliver a second, more impactful shock: "We've built a whole lot more," Hou said. "There's a lot more economic value and assets that are at stake."
For China, Kimi K3 is a triumph in the face of restrictions on the latest Nvidia (NVDA) chips, EUV equipment and other U.S. export controls. The release triggered an outpouring of national pride on Chinese social media. In a post shared on the platform QQ, one blogger called K3 a "source of our glory" in Chinese, saying "this level of quality can't be achieved through distillation," referring to a machine learning technique used to transfer a larger, more powerful model's capabilities to a smaller one.
Moonshot, along with other Chinese AI labs DeepSeek, MiniMax (HK:100) and Alibaba $(BABA)$, have found themselves at the center of a distillation controversy. Anthropic has accused them of engaging in industrial-scale distillation attacks, using thousands of unauthorized accounts and exchanges to replicate Anthropic's reasoning chains to train their own models for a fraction of the cost. In January, the release of Kimi K2.5 caused a stir when users on social media noted the model would sometimes refer to itself as Anthropic's Claude - a likely signal that the model was trained on data from Anthropic.
However, Kimi's latest model received praise for its innovation. "There are genuine algorithmic innovations in the model," Silicon Data's Hou said. In an X post, NYU professor and AI researcher Ravid Shwartz-Ziv pointed to Kimi's hybrid linear attention software, which allows the model to process large quantities of data with less memory consumption.
Concerns of distillation are largely overblown, according to Florian Brand, research engineer at AI company Prime Intellect. Distillation is a common practice that occurs during a small window of the training process, he told MarketWatch.
U.S. labs have utilized this practice as well. Earlier this week, AI lab Thinking Machines announced its latest Inkling model, which used outputs generated by Kimi K2.5 during the fine-tuning phase. Cursor's Composer 2 model, launched in March, was pretrained on Kimi K2.5.
Open models lower costs and give users more options, Brand said. "Open source drives innovation forward," he told MarketWatch. "Without them, we would just have closed labs deciding who gets access to what technology."
However, some argued the reaction to Kimi K3 was overhyped. Max Weinbach, analyst at Creative Strategies, told MarketWatch that K3 is a "good model" but that OpenAI's GPT-5.6 Luna is "still better and still the cheapest one."
"The GPT 5.6 series tends to ground itself in real data when doing knowledge work. It's better at citing where that data came from. It's a little more reliable on the hallucination front," he said.
Indeed, Moonshot included the following caveat in its announcement of the model: "Despite being a highly competitive model overall, K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol."
Some on social media argued that Kimi K3 and other Chinese models were "benchmaxxing," meaning that the models are trained to perform specific tasks to score well on standardized tests but fall short in general usage areas.
Weinbach also said that Kimi K3 "doesn't really change anything" in terms of its compute economics, adding that the conception that open models are cheap to run isn't necessarily true.
"Because this model is so large, you need GB200 or GB300 to run it well and those are $4 million to $6 million per rack," he explained, referring to Nvidia's Blackwell architecture. "This is not affordable for a company to just go buy and run."
-Christine Ji -Hannah Pedone
This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.
(END) Dow Jones Newswires
July 17, 2026 18:31 ET (22:31 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments