Moonshot AI unveils Kimi K3, aims to close gap with US rivals

Moonshot AI unveiled Kimi K3 as a 2.8 trillion-parameter open-weight model with a 1 million-token context window, a scale the Beijing startup said makes it the world’s largest open AI system. Moonshot said Kimi K3 performs competitively with leading proprietary models from Anthropic and OpenAI, while still trailing their best systems, putting the release squarely in the race to narrow China’s gap with U.S. frontier labs.

The launch extends a rapid climb for Moonshot, which was founded in early 2023 and built the Kimi chatbot family around large models aimed at coding, knowledge work and agentic tasks. The company has already used Kimi K2 to claim another milestone, describing it as the first open model to scale pre-training to 1 trillion parameters. Founder and chief executive Zhilin Yang, a Tsinghua University graduate who earned his doctorate from Carnegie Mellon University in 2019, has become one of the most closely watched figures in China’s AI sector.

Market signals around the company suggest the competition is not only about benchmark bragging rights. South China Morning Post said Moonshot’s Kimi and Zhipu’s ChatGLM together drew nearly 35 million monthly active users, a sign that Chinese AI products are starting to build real consumer and developer audiences at home. Reuters has also reported Moonshot at about $31.5 billion in valuation amid an ongoing funding round, while other 2026 reports said the company raised roughly $2 billion and was seeking a valuation as high as $30 billion.

The open-weight release matters because it changes the terms of adoption. A model that can be downloaded, adapted and deployed more freely can pull in developers who want to control costs, customize workflows or avoid dependence on a closed API. That puts pressure on U.S. model makers to justify premium pricing and stronger performance claims, while giving Chinese labs a route to spread their systems beyond domestic chat applications into enterprise software and agent-driven tools.

It also sharpens the policy debate around AI competition and export controls. A model released openly from Beijing can circulate through global developer communities even if the company itself remains subject to capital, chip and trade constraints. Moonshot’s latest move shows that China’s AI contest with the United States is now being fought not just on raw capability, but on distribution, openness and the ability to turn technical progress into adoption.