On July 16, Moonshot AI, a Beijing-based AI lab, released Kimi K3, the largest open-source model ever built, at 2.8 trillion parameters, rivalling top American AI models. Within a day, Wall Street and Silicon Valley were treating it as a genuine rival to models from OpenAI and Anthropic, and Chinese AI stocks moved sharply on the news. This episode shows that the AI frontier is no longer a two-company story.
Event Context
Kimi K3 arrives after 18 months in which Moonshot AI’s market position had eroded following DeepSeek’s rapid rise. The comeback rests on real architectural work, not just scale. K3 is built on two innovations called Kimi Delta Attention and Attention Residuals, designed to help information flow through longer sequences and deeper models, paired with a Mixture-of-Experts design with 896 experts, of which only 16 activate per token — so a 2.8-trillion-parameter model behaves, computationally, more like a much smaller one.
The architect behind it, Yang Zhilin, has built a career almost entirely inside the American research pipeline before he ever ran a Chinese lab. He joined Carnegie Mellon University to pursue a PhD. While at CMU, he interned at Google Brain and Meta, and co-authored two influential transformer papers before GPT-3. He then returned to China and contributed to Huawei’s PanGu model and the Beijing Academy of Artificial Intelligence’s Wu Dao project before co-founding Moonshot in March 2023 with Tsinghua schoolmates Zhou Xinyu and Wu Yuxin.
His own operating principle, as he has put it, favours scale over cleverness: new algorithms exist mainly to make scaling work better, not to substitute for it.
The American labs’ focus on closed, tightly controlled systems has coincided with a period in which one of them, Anthropic, was actively fighting its own government over access to its best models.
On June 12,, the U.S. government ordered Anthropic to suspend foreign nationals’ access to Fable 5 and Mythos 5 under an export control directive issued on national security grounds — a directive that forced a complete shutdown rather than just blocking users outside the U.S.
As a result Chinese AI usage hit 25 trillion tokens in a single week, 78% above U.S. volumes, as at least one startup moved all its traffic to DeepSeek, citing savings of millions of dollars. In the same three-week stretch, Washington also put OpenAI’s GPT-5.6 Sol behind customer-by-customer government approval.
The government then changed course as the Department of Commerce lifted the export controls on Fable 5 and Mythos 5, and Anthropic said it would begin restoring access, with Commerce Secretary Howard Lutnick saying the agency had worked with Anthropic to align the models across the government.
Team Analysis
Moonshot says the model substantially outperforms Anthropic’s Opus 4.8 in certain parameters but concedes K3 still trails Fable 5 and GPT-5.6 Sol in overall performance.
Match Outlook
Once the weights are public, the open-source community is expected to begin distilling, quantising and optimising the model almost immediately — the same pattern that followed Llama, Qwen and DeepSeek, happening faster with each cycle.
That prospect is already reordering enterprise assumptions. If K3’s benchmark claims hold up once outside researchers can test the open weights, it will be difficult for closed-model providers to justify premium pricing on capability alone.
It has also unsettled Beijing as China’s Ministry of Commerce is reportedly weighing export controls of its own on open-weight AI models, worried that freely downloadable Chinese systems are ending up in Western hands, even as it keeps API access open.
Whatever happens next, the contest over the next generation of models built on top of Kimi K3’s weights — in the U.S., China, and everywhere in between — is no longer a two-horse race.
source
But the underlying dynamic had already shifted as Chinese open-source models were reporting to be 60% to 90% cheaper than leading Anthropic and OpenAI offerings, and enterprises increasingly began routing routine tasks to whichever model was “cheapest that’s good enough” rather than defaulting to the U.S. frontier leaders.
For now, K3 is only accessible as a hosted service so people can try and an API businesses can call. But they cannot download the model. That is slated to change on July 27, when Moonshot has committed to releasing K3’s full weights, allowing organisations to run the 2.8-trillion-parameter model entirely on-premises.
