The AI industry is not quite experiencing a DeepSeek 2.0 moment, but it feels very close. The leading Chinese AI labs have been on a roll lately, releasing a series of almost cutting-edge open-source models. Z.ai released GLM 5.2 in June, Moonshot AI released Kimi K3 last week, and Alibaba released Qwen 3.8 this Monday.
Silicon Valley and Washington started talking about the models immediately, especially K3, which is widely seen as the best of the bunch. David Sacks, a venture capitalist and AI adviser to President Donald Trump, called the performance of Moonshot’s model “concerning.” Earlier this week, Commerce Secretary Scott Bessent suggested the US might impose sanctions on Chinese AI companies.
On Wednesday, Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged that the Trump administration has “information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model,” which he said amounted to “stealing proprietary US technology and undermining American research” and was “unacceptable.” (Moonshot AI did not immediately respond to a request for comment.)
The new Chinese models have a few things in common: Third-party benchmarks show that they perform nearly as well as the best Western models; they are optimized for agentic coding tasks (the hottest thing in AI this year); and they are or will soon be released with open weights, making them accessible and transparent.
But perhaps the biggest parallel between the current moment and January 2025—when the world was shocked by DeepSeek’s R1 model—is that it reaffirms how American and Chinese AI labs are taking diverging paths when it comes to being open or closed.
When it first burst onto the scene, DeepSeek challenged the premise that only closed-source models built with billions of dollars of investment in compute infrastructure and training could achieve frontier performance. But since then, Western AI labs have continued developing AI the same way, and now American frontier models feel more roped-off than they were a year ago.
Anthropic said for months that its latest Mythos model was so dangerously good at hacking that only approved collaborators could use it. When it was finally released more widely, the White House responded by issuing broad export controls, which forced Anthropic to take Mythos and its less capable sister model, Fable 5, offline temporarily. OpenAI similarly delayed the release of GPT 5.6 after it received a request from the White House.
In China, meanwhile, the situation looks very different. Chinese startups and tech giants have doubled down on open source: Anyone with a good enough computer environment can now download an open-weight model, run it locally, add customizations, and overall enjoy a much greater degree of freedom than OpenAI and Anthropic would ever allow. In many ways, the open versus closed debate is more entangled with the US versus China debate than ever before.
There are a lot of reasons why Chinese labs have chosen a business strategy built atop open-source models. Being the newer, smaller fish in the AI field, making their models free and open can help Chinese firms attract more users, collaborators, and media spotlight. It also puts them in a separate lane of competition from the one that OpenAI, Anthropic, Google, SpaceX, and other deep-pocketed giants are in.
Earlier this year, rumors spread that Alibaba might be considering joining the closed-source race after it rearranged its corporate AI model development teams. But the tech giant announced on Monday that it would again release the latest version of Qwen—its line of open-source models beloved by the global tech community—with open weights, signaling to customers and the public it is not pivoting away yet.
Perhaps the strongest validation of the Chinese approach is the models themselves. Chinese labs have produced what are widely considered the world's best open-source AI models, showing that American companies no longer have an exclusive hold on building the most capable systems.
Arena AI, a crowdsourced model evaluation platform, now ranks K3 as the best model when it comes to web development tasks and number four in agentic tasks, just below Anthropic’s Fable and Opus 4.8, as well as OpenAI’s GPT 5.6. Artificial Analysis, an independent AI benchmarking company, ranks K3 in the third spot in its intelligence index.
Shortly after Moonshot AI released a preview version of K3 on July 16, people around the world began rushing to try it out, consuming so much inference computing resources that the company is temporarily restricting new users from signing up.
As more people start using open-source Chinese models with capabilities nearly on par with that of their Western competitors, some have started questioning whether paying for OpenAI or Anthropic’s offerings is really worth it. The fact that multiple Chinese labs were able to create capable agentic models and were unafraid to release them to the public is poking holes in the popular understanding that OpenAI and Anthropic are miles ahead of the competition.
Just like with DeepSeek, K3 fans are now celebrating it as evidence that Western AI models are overhyped and overprotected. “It is absolutely wild how much love Kimi got,” Rui Ma, founder of the independent research firm Tech Buzz China, said in a social media post, referring to the company’s announcement that demand for K3 had overwhelmed its servers. “Only made possible by the poor comms and decisions from [Silicon Valley] labs in the past year,” she added.
“I think Anthropic has overhyped the risks, or described risks that are coming soon but do not currently proliferate,” says Nathan Lambert, an independent AI researcher in Seattle who recently visited Moonshot AI’s office in China. He admits, though, that just like the rest of the general public, he has little firsthand information about how Mythos really performs. “We rely on a few private companies and a federal government with depleted state capacity to make that judgment call,” he says.
In addition to challenging the mainstream Western narrative about AI, Chinese open-source models are also proving to be a practical commercial replacement for American closed models. They aren’t just talked about on X and benchmarked against other models—they have become a popular and readily usable option for a growing number of Western startups and individual users.
“There's a real shift toward actually using these newest models that started with GLM 5.2,” Lambert says, referring to Z.ai’s model. “Even weeks after the GLM 5.2 release, I hear from AI researchers in the Bay Area that they are still using it for a core part of their workflow. Kimi, being a stronger model, will only do more, especially in areas like cybersecurity, where Mythos, Fable, and GPT 5.6 are effectively unusable.”
In fact, this is already happening. On Tuesday, OpenAI disclosed a concerning incident in which its GPT-5.6 Sol model hacked into the production system of Hugging Face, an open-source AI platform. Hugging Face has said it resorted to using the open-source model GLM 5.2 to analyze the cyberattack because other frontier models were refusing to help due to built-in safety guardrails.
Chinese AI models also tend to be cheaper than Western alternatives, but that’s not necessarily their main selling point. While models like K3 charge less for tokens, early tests show that they may require more tokens than Western models to solve the same problems, thus making the cost gap smaller. “In my fairly limited use, [K3] also seemed very token-hungry. It's not obvious to me that this model is actually that cheap to run,” wrote Dean Ball, a former White House AI adviser who recently joined OpenAI as its head of strategic futures, in a social media post in which he also praised K3 as “a very good model.”
Even if they are relatively token-hungry, models like K3 do ultimately challenge a fundamental assumption that has driven OpenAI’s and Anthropic’s strategy for years: that AI labs need infinite funding to scale up their compute capabilities just to make better models. “In the end, open-weight models deter further AI capex,” Ball wrote.
Instagram Engineering
Shopify Engineering