What is an open-weight AI model, and why is China giving its best ones away for free

In the world of artificial intelligence, a model's "weights" refer to the full set of mathematical parameters it learned during training. Put simply, these weights can be thought of as a model's "brain" — they contain everything the model has learned about how to understand text and how to respond to a question.

In "closed" models — the kind offered commercially by many leading American companies — these weights are stored on the company's own servers, and users can only access the model through an API. In "open-weight" models, by contrast, these parameters are published publicly, allowing anyone to download and run the model on their own infrastructure, or even modify it.

Chinese AI startup Moonshot AI's Kimi K3 model, announced last week, reignited exactly this debate. The company claims the model outperforms leading American systems on some tasks — and says it achieved this at a fraction of the usual training cost.

But what has truly unsettled Silicon Valley isn't the model's performance so much as Moonshot's decision to release its weights for free. That means developers, researchers and companies around the world can access a leading AI system without paying any subscription fee at all.

Experts believe there are both technical and geopolitical motivations behind this strategy. On one hand, having a broad developer community test and improve an open-source model can accelerate a company's own research process. On the other, the move is also seen as a way for China to showcase its technical capability in AI on a global stage.

The spread of open-weight models poses a serious threat to the business models of American companies. Closed-model providers generate much of their revenue from API usage fees, and the existence of a free, capable alternative directly undermines that revenue model.

Some American companies are trying to respond to this trend by releasing their own open models. But industry officials stress that an open-source strategy isn't the right approach for every company, since developing such models still requires substantial cost.

Experts note that open-weight models also bring a set of security concerns. Once a model is released publicly, preventing malicious use becomes far harder, since the company loses any ability to control how the model is used. This has sparked new debates among regulators.

Open models are also raising national security concerns. Some policymakers argue that letting advanced AI capabilities circulate freely could increase the risk of the technology being misused.

The debate points to one of the fundamental tensions the AI industry is likely to face in the period ahead: the balance between openness and control. Closed models offer greater security and revenue control, while open models accelerate innovation and democratise access to the technology. Which approach will ultimately dominate remains uncertain.