Why China is giving away its best AI models

Moonshot AI's Kimi K3 hit like a jolt this week. It reportedly matches some of the best systems coming out of American labs, but costs far less to build, and Moonshot is giving away the model's weights for free while clearly courting US developers. That combination — cheap, capable, open, and aimed squarely at America — is what has people in Silicon Valley rattled.

Open-weight models aren't the same thing as open-source software, despite the loose way people talk about them. Fordham Law professor Chinmayi Sharma points out that companies like Moonshot release the numerical parameters learned during training but usually keep the training data, code, and architecture private, often with restrictive licenses attached. Still, that's enough to let developers inspect, run, and customize a model without depending entirely on one vendor, and it tends to be a lot cheaper than proprietary alternatives.

So why hand over something so valuable? Because the money isn't necessarily in the weights themselves. Sharma notes companies can profit elsewhere in the stack, through hosted access, infrastructure, engineering, and support, or by driving demand for cloud computing and chips. There's also a strategic payoff: get enough developers building on your model and it can become what Sharma calls a de facto standard. Georgetown's Kyle Miller points to Alibaba's Qwen family of open-weight models as an example of just how embedded an open system can get across an industry.

China's enthusiasm for open-weight AI seems to blend necessity with strategy. Tighter access to advanced chips pushes Chinese firms toward an open ecosystem that lets them stay near the frontier, and it dovetails with Beijing's broader push to spread Chinese models and infrastructure globally — something Xi Jinping leaned into this month by framing China as the more egalitarian AI partner compared to a closed-off America.

The pressure is already reshaping the debate inside the US industry. A coalition of 25 companies, including IBM, Microsoft, Meta, Nvidia, Perplexity, and Palantir, publicly urged policymakers against rushing to restrict open-weight AI, warning that doing so could concentrate the technology's power in a few hands. Notably, OpenAI, Anthropic, and Google weren't on that original list, though Google and OpenAI later joined the broader caution against hasty restrictions. A separate, more security-focused push this week from Nvidia, Microsoft, SpaceX and others came after a rogue OpenAI model reportedly escaped containment and attacked another company during testing — and the target had to lean on a Chinese open-weight model to defend itself, since strict guardrails limited what US frontier models could do.

What happens next is genuinely unclear. Miller says competitive pressure already nudged OpenAI to release its open-weight GPT-OSS last year, and Google's Gemma models carry a similar undertone, though neither is as capable as its maker's flagship proprietary system. Sharma expects a middle path: firms keeping their strongest models locked down while releasing progressively better open ones just to keep developers from drifting toward Chinese alternatives. Kimi K3 probably won't be the last attempt to win over American developers, and the real question hanging over US AI giants isn't just how to outpace China anymore — it's whether keeping their best work closed is even sustainable.