Kimi K3: The Chinese Breakthrough That Worries OpenAI and Anthropic

Understanding Open Weight Models

When an artificial intelligence model is trained, it absorbs a vast amount of data. The information it extracts translates into numerical values known as "weights." These weights are essential because they determine how the model reacts to various stimuli, whether it's answering a question, writing text, or solving a complex problem. Making these weights publicly accessible allows anyone to use them to run the model on their own infrastructure, adapt it to specific needs, or study it to better understand its internal workings.

However, it is crucial to differentiate an open weight model from open source software. Open source software reveals its entire code, allowing anyone to use and modify it freely. In contrast, an open weight model only shares part of its components. While the weights may be made public, the training data and details of the learning process often remain confidential. Thus, the openness is never complete.

On the other hand, closed models, such as Claude, ChatGPT, or Gemini, reveal nothing of their internal composition. The creators of these models strictly control their use, leaving little room for users.

This partial openness of open weight models has concrete implications for digital companies. These models are often less expensive to operate, can be hosted on internal servers rather than with third-party providers, and easily adapt to specific uses. Start-ups like Cursor are betting on this strategy, using Chinese open weight models such as Kimi, GLM, or DeepSeek, rather than the closed models from American giants.

China Leading in Open Weight Models

Before the emergence of DeepSeek, the last cutting-edge open weight model dates back to 2019 with OpenAI's GPT-2. The introduction of R1 in January 2025 marks a significant return of open models after several years of absence. It is also the first time such a model comes from China.

China has developed a clear strategy around openness, politically supported. On July 17, 2026, during the World AI Conference in Shanghai, President Xi Jinping stated: “We often say in China that a single string cannot make music, and a single tree does not make a forest […] The development of AI should not be a solo performance by one country, but a symphony of international cooperation.”

After DeepSeek, Moonshot AI launched Kimi K3 in mid-July 2026. This model fits into an already rich ecosystem that includes GLM from Zhipu AI (now Z.ai) and Qwen from Alibaba. This strategy seems fruitful. On OpenRouter, a popular platform among developers for accessing various AI models, the share of traffic to Chinese models increased from 20% to 48% between June 2025 and June 2026, while the share of American models dropped from 74% to 32% during the same period.

However, contrary to the statements of the General Secretary of the Chinese Communist Party, support for open weight does not rest solely on principles. For example, Moonshot charges up to $15 per million tokens for the use of its API for Kimi K3, compared to $4 for the previous version. Although the weights are public, any company wishing to commercially exploit the model must sign a separate agreement with Moonshot. Thus, openness comes with a cost and conditions.

Moreover, cybersecurity concerns are prompting the Chinese government to consider restricting access to its most powerful models. This reasoning is similar to that of Washington, which a month earlier had limited access to Fable 5 and Mythos 5.

The Stakes for Nvidia, OpenAI, and Anthropic

In the face of the rise of Chinese models, the Trump administration is considering restrictions specifically targeting open weight models from China. On July 24, 2026, an open letter initiated by Nvidia and signed by over 70 companies, including Microsoft, Meta, IBM, Dell, Palantir, and later OpenAI, calls on authorities to maintain access to these models.

In both the United States and China, these positions are driven more by economic interests than by genuine ideological visions.

For Nvidia, the motivations are clear. The company sells chips, not models. The more open weight models there are to operate, the more companies need hardware to host and run them. Nvidia is already using the argument of openness to promote its GPUs in Europe and has even created the Open Secure AI Alliance to advocate for open tools that security teams can inspect and adapt.

For OpenAI, the position is more nuanced. The company develops closed models, and one of its executives, Dean Ball, has expressed the need to find a balance between openness and security.