When Washington restricted exports of advanced Nvidia chips to China in 2022, the policy’s stated goal was to slow the development of Chinese artificial intelligence. Four years later, Chinese AI labs have not only kept pace but are distributing their best models freely to the world, a strategy that turns export controls into a distribution amplifier.
The latest and most dramatic example arrived July 27, when Moonshot AI released the weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with 896 experts, 16 active per token, and a context window of one million tokens. It is the largest open-weight model ever published. Any developer can download it, modify it, run it on their own hardware, or incorporate it into a commercial product.
Kimi K3 is not an isolated case. Alibaba’s Qwen 3.8 Max, a 2.4-trillion-parameter model that the company claims performs behind only Claude Fable 5, entered preview this month. Z.ai (formerly Zhipu) released GLM-5.2 under the MIT license on Hugging Face. Tencent’s Hunyuan Hy3 shipped under Apache 2.0. Baidu, which long kept its Ernie model behind an API, reversed course and published weights.
The numbers on adoption are striking. Qwen has surpassed one billion cumulative downloads on Hugging Face. By March 2026, it accounted for more than half of all open-source model downloads globally, with more user-generated variants than Google and Meta combined. A study by MIT and Hugging Face found that Chinese open-weight models represented 17.1 percent of global AI model downloads in the year ending August 2025, exceeding the US share of 15.86 percent for the first time.
The strategic logic behind the giveaway is the inverse of the usual Silicon Valley playbook. American frontier labs like OpenAI and Anthropic keep their most capable models behind paid APIs, monetizing access per token. Chinese labs lack the same abundance of advanced hardware for serving inference at scale, so they release weights and let others host and serve the model on their own infrastructure. Databricks, cloud providers, and enterprise data centers become distribution nodes.
Kyle Chan of the Brookings Institution described the effect as unlocking compute capacity that other organizations have already invested in and built up. By open-weighting a model, the developer turns the global installed base of GPU clusters into its delivery network, bypassing the need to build its own inference infrastructure.
The implications extend beyond the digital realm. A March 2026 report from the US-China Economic and Security Review Commission identified two reinforcing feedback loops in Chinese AI strategy. The digital loop works through open releases, developer adoption, community iteration, and model improvement. The physical loop operates through deployment in manufacturing, logistics, and robotics, generating proprietary real-world data that feeds back into model improvements.
Export controls targeting training hardware address only the digital loop’s upstream. They leave the physical loop untouched, because deployment-side data accumulation on China’s industrial base does not require frontier chips. The commission concluded that the battlefield has moved beyond the laboratory to the factory floor.
Adoption data suggests the strategy is working inside the United States as well. An analysis by AlphaMatch found that roughly 80 percent of US AI startups now integrate at least one Chinese open-source model. Chinese models account for approximately 60 percent of tokens consumed by US companies on the OpenRouter inference platform. A developer who builds an application on Qwen or DeepSeek is unlikely to switch back when the next-generation proprietary API raises its prices or changes its terms.
The open-weight approach does not foreclose revenue. Moonshot charges $3 per million input tokens and $15 per million output tokens for hosted API access to Kimi K3, and offers enterprise subscriptions. But the primary value is strategic, not transactional. By distributing weights freely, Chinese labs build global dependency on their model ecosystems, creating a structural counterweight to the dominance of US API providers.
That dominance has not gone unchallenged in Washington. The Trump administration has revived discussions around banning Chinese AI models from government systems, and OpenAI’s head of strategic futures described the scenario of widespread open-weight availability in stark terms. But the tools are already downloaded, running on servers worldwide, and deeply integrated into the development workflows of the next generation of AI applications.
Sources: Why China is giving away its best AI models (The Verge, Jul 27); How the Two Loops of Chinese AI Competition Work (USCC Report, Mar 2026); Kimi K3 is the largest open-weight model yet (Tom’s Hardware, Jul 27); China open-source AI models surpass US in global downloads (MIT Technology Review, 2026)