What to know about Moonshot AI and its new open weight model Kimi K3
The release of the Chinese AI model Kimi K3 was a flashpoint in the AI world, sharpening the debate over whether the most capable models should be closely held by individual companies, usually U.S. tech firms, or freely and transparently distributed in the tradition of open-source software.
Not only did the high-performing Kimi K3 challenge the cherished Silicon Valley notion that Western AI labs still lead their Chinese counterparts in large language models, though likely by only a hair, but it also raised the possibility that more transparent open-weight models may ultimately be easier to control, and therefore safer, than closed-weight models whose makers must guard against misuse.
What is Kimi K3?
Beijing-based Moonshot AI released Kimi 3 in mid-July, followed by the model’s full weights on July 27. With roughly 2.8 trillion parameters, it is likely the largest open-weight model ever released. The model is natively multimodal, meaning it can process images and audio alongside text. It can also reason across a 1-million-token context window.
Why did Kimi K3 take off so quickly?
The Kimi K3 weights are available on Hugging Face, a widely used clearinghouse for open-weight models. Hugging Face CEO Clem Delangue said the model reached No. 1 on the platform’s trending chart with more than 4,000 likes within 30 minutes of its release, the fastest growth the site had ever recorded. By the next day, the Kimi K3 repository had about 7,700 likes and tens of thousands of downloads.
Many developers saw Kimi K3 as a tipping point: a very large open-weight model that could compete with the biggest closed, pay-per-token models. Moonshot reported frontier-level performance on coding, agentic, reasoning, and knowledge tasks, placing behind, but not far behind, Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol in overall evaluations. Kimi K3 also outperformed many frontier models on widely used coding benchmarks.
The model excels at long, complex coding, agentic, and multimodal tasks. In short, Kimi K3 gave developers a chance to modify, fine-tune, and build applications on top of a highly capable model at a fraction of the cost of using a closed one.
Who is behind Moonshot AI?
Moonshot AI was founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, who had been classmates at Tsinghua University in China. Yang, the company’s CEO, earned a PhD from Carnegie Mellon University, home to one of the United States’ leading computer science programs. While completing his studies, he interned at Google Brain and Facebook AI Research, or FAIR, contributing to major natural language processing papers at both companies. His PhD adviser, Ruslan Salakhutdinov, said Yang was determined to return to China and start his own company rather than remain in the U.S.
What is the company’s history before Kimi K3?
Moonshot’s original Kimi chatbot attracted attention in China in October 2023 for its ability to process long stretches of Chinese-language text. The company continued refining the model before unveiling Kimi K3 in July 2026, first through its API and platform and later as an open-weight release.
How much is Moonshot AI worth?
In May, Moonshot AI raised $2 billion in a round led by Long-Z Investments, Meituan’s venture arm, with participation from China Mobile. Afterward, the company was valued at $20 billion. While Moonshot’s total funding to date is a matter of speculation, the South China Morning Post reported in early May that the company had raised $3.9 billion over the past six months.
Why did Moonshot release the weights publicly?
Because the U.S. restricts Nvidia and other companies from selling their most powerful AI chips to China, Chinese AI labs often struggle to secure enough computing power to serve their models through the cloud. Demand for Kimi K3’s hosted version maxed out Moonshot’s compute resources within days of launch, forcing the company to temporarily pause new consumer subscriptions. That helps explain why Chinese model makers often release models for free with their weights publicly available, allowing developers, researchers, and third-party hosts to run them on their own infrastructure. The approach expands the model’s reach without requiring Moonshot to build enough serving capacity to meet global demand.
How has Washington reacted?
The U.S. already uses trade restrictions to keep the most powerful AI chips out of the hands of Chinese researchers. After Kimi K3’s release, reports emerged that OpenAI and Anthropic had lobbied the government to ban Chinese open-weight models from use by U.S. developers and businesses. The argument was that such models could contain difficult-to-detect features that make them susceptible to malicious uses, including cyberattacks on critical infrastructure or the development of bioweapons. U.S. officials have also raised concerns that Chinese AI labs may be using outputs from closed U.S. models to train their own systems. Anthropic, for instance, alleges that Moonshot AI used outputs from Anthropic models to train Kimi models.
However, tech companies have pushed back against a U.S. ban on Chinese open-weight models. Every major U.S. AI lab except Anthropic signed an open letter last week written by Nvidia CEO Jensen Huang arguing that broad restrictions would harm U.S. startups, universities, and other organizations that rely on accessible models. “Open weights let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else,” Huang wrote. Three days later, on July 27, Nvidia launched the Open Secure AI Alliance with 44 founding members.
On the matter of safety, Huang argues that banning open-weight models would deprive U.S. organizations of a valuable tool for defending against cyberattacks while doing little to prevent foreign adversaries from using the same models offensively. In a Monday blog post, Anthropic CEO Dario Amodei agreed with that idea and denied that his company had called for a U.S. ban on open-weight models. Instead, he argued for maintaining restrictions on advanced AI chip sales to China, cracking down on Chinese labs that use U.S. model outputs for training, and requiring frontier models to undergo rigorous safety testing before release.
1 month ago
