A Chinese Open Model Just Matched OpenAI and Anthropic

In Brief

Chinese AI lab Moonshot AI released Kimi K3, an open-weight model that performs on par with Anthropic's Opus 4.8 and OpenAI's GPT-5.5 on many benchmarks. 

Washington is considering sanctions and a possible ban on Chinese models, while nearly 200 startups warn a ban would  kill hundreds of companies overnight.

What Happened

Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that performs on par with top proprietary models, including Anthropic's Opus 4.8 and OpenAI's GPT-5.5, across many benchmarks. Parameters are the internal settings that determine how capable a model is, and open weight means anyone can download those settings and run the model on their own hardware. K3 can also process images and has a one-million-token context window. Tokens are the amount of text a model can handle at one time. Demand was so intense that Moonshot paused new subscriptions days after launch, and Vercel CEO Guillermo Rauch said it was the first time an open model beat every proprietary one on his company's web engineering benchmark. Thinking Machines released its own open-weight mode called Inkling the same week, and Alibaba's Qwen team says its new 2.4-trillion-parameter Qwen 3.8 will go open weight soon.

The launch unsettled leaders in Washington. White House science policy chief Michael Kratsios said the administration has information that Moonshot distilled Anthropic's Fable model to develop K3. Distillation means training your own model on another model's outputs, effectively copying its capabilities. Kratsios said Moonshot did this at large scale while evading detection, and acquired servers with Nvidia GB300 chips despite a ban on their sale to Chinese entities. Treasury Secretary Scott Bessent promised an investigation, warned that "open source is not open season on American IP," and put sanctions on the table. Axios reports the administration could ban cutting-edge Chinese models entirely.

Nearly 200 companies, including Y Combinator, formed the Little Tech Association to urge the administration not to cut off the Chinese open-weight models many startups depend on. One founder told Politico "there will be hundreds of companies that instantly die." A White House official called reports of a coming ban "baseless speculation."

SmarterX founder and CEO Paul Roetzer discusssed the standoff, and the difference between open weight and open source, on Episode 226 of The Artificial Intelligence Show.

The Key Numbers

2.8 trillion - Parameters in Moonshot AI's Kimi K3

1 million  - Kimi K3's tokens/context window

~200 - Companies in the Little Tech Association opposing banning open models

90% - Inference margins Gavin Baker says a two-to-three-lab world would hand the winners

2.4 trillion - Parameters in Alibaba's Qwen 3.8, the next model set to go open weight

Why Open Weight Is Not Open Source

The terms get confused, even by tech leaders. Roetzer's analogy: a closed model such as ChatGPT or Claude is renting a car. You can drive it, but you cannot modify it, and the company can take it back. Open weight is owning the car. You can rebuild the engine, paint it, rent it out, but you never get the engineering drawings. Open source is the car plus the CAD files, manufacturing specifications, and assembly instructions. With open weights, you never see the training data or methods. "That's the secret sauce that they're not giving you," says Roetzer. "Open source, they give you everything, including the secret sauce." Kimi K3, like almost every major "open" model, is open weight, not open source.

The reaction depends on who is reacting. Investor Gavin Baker wrote that "Kimi K3 may be an important inflection point for AI, potentially negative for Anthropic and OpenAI, while being net positive for essentially every other company in the world," though "the real Sputnik moment would be an open source frontier model." A world where two to three labs earn 90% margins on inference, the part where a model actually answers questions, is in his view bad for every other layer of the economy. Box CEO Aaron Levie agrees that cheaper AI benefits the whole ecosystem. Roetzer's caution: Levie builds on models he does not make, and Baker invests in companies cheap AI helps. "No one's neutral in this, is what I'm saying," he says.

Policy insiders doubt a ban is coming. Dean Ball, an architect of the administration's original AI policy, called K3 "a very good model" whose performance cannot be explained by distillation alone, and argued open-weight models are "effectively ungovernable." His prediction: The administration will create regulatory risk around Chinese models rather than ban them. Former White House AI czar David Sacks called K3's rise concerning, then aimed his fire at U.S. regulation instead, arguing "permissionless innovation is how America won the internet."

The distillation complaint is hypocritical. Roetzer thinks the accusation is likely true, but a losing argument because every major model was trained on intellectual property that did not belong to the labs that built it.

"You have companies that stole to create models complaining that someone else is stealing their models.

They're never going to win the public battle. That is done. So good luck arguing that one in the public."

— Paul Roetzer, founder and CEO of SmarterX, Episode 226 of The Artificial Intelligence Show

SmarterX Take

Some form of government response is coming. "I definitely think they're going to take some action. I don't know what it is," says Roetzer. The harder question is why China gives frontier-grade models away at all. Roetzer pointed to former Australian Prime Minister Kevin Rudd, who knows Xi Jinping personally, speaking on a recent Ezra Klein show episode. Rudd describes a leadership that sees dominance as China's predetermined place in history, pours everything into manufacturing, high technology, and control of supply chains, and asks citizens to "eat bitterness" through the painful stretch.

Roetzer's takeaway: China is willing to go to lengths America may not be, and flooding the market with cheap, capable models is one of them. A frontier model given away for free is a strange business decision but a coherent state strategy to undercut the economics of American AI. Leaders choosing which models to build on should know that cheap intelligence is no accident.

What to Watch

Regulatory risk, not an outright ban, is the most likely next move. Ball predicts the administration will make Chinese models legally and commercially risky to use rather than prohibit them. For startups built on those models, the uncertainty must be factored into decisions.

The weights themselves are the point of no return. Moonshot says it plans to publicly release K3's full model weights. That is the point behind Ball's "effectively ungovernable" warning: once weights are downloaded around the world, they cannot be taken back. 

Which AI Models Do Companies Actually Provide?

ChatGPT is the most widely provided AI platform in business, with 59% of professionals saying their organization offers it, according to the 2026 State of AI for Business Report. Microsoft Copilot (46%), Google Gemini (42%), and Anthropic Claude (37%) follow. Today's corporate AI stack runs almost entirely on the closed U.S. models at the center of this fight, which is why a wave of cheap, near-frontier open models could reshuffle it.

The full report, based on 2,100+ responses from professionals across roles, functions, and industries, covers which tools organizations provide, where they stand on adoption, and what is holding them back. Read the full report →