The Open Weights Fight Just Swapped Sides

The largest open weight model in the world is Chinese, and Beijing has been holding meetings about whether to let the rest of us have it.

Three things happened inside nineteen days in July. They are only strange when you read them separately.

On July 24, twenty five US companies including Nvidia, Microsoft, Meta, IBM, Dell, Palantir and Hugging Face sent a policy letter to Washington urging the government not to impose premature restrictions on open weight AI models. The stated case was competition, security and US AI leadership.

On July 27, Moonshot AI published the weights for Kimi K3, a 2.8 trillion parameter model with a one million token context window, on Hugging Face and GitHub. It is likely the largest open weight model released to date. It is not American.

And on July 8, Fortune and Reuters reported that Chinese ministries had been meeting Alibaba, ByteDance and Zhipu AI about restricting overseas access to China's most advanced models, along with new limits on who can fund domestic AI startups.

## The argument stopped being about values

For two years, open weights was debated in the United States as a question of principle. Does releasing model weights accelerate research or accelerate misuse? Both sides had papers.

That framing is now behind the facts. Open weights has become a supply question, and the two sides are drifting toward each other's old positions.

American industry is lobbying to keep open release legal in part because a restriction now would not stop open frontier models. It would only stop American ones. The counterfactual is a market where the best openly available model is governed by a jurisdiction Washington does not regulate.

Beijing is moving the other way for the mirror reason. Open release has been the most effective distribution channel Chinese labs ever had. It put Chinese models into Western startups, benchmarks and tutorials at zero marketing cost. Once you are the one with something worth withholding, free distribution starts to look like a concession.

There is a tell in the chip file too. On July 9, Beijing signalled it would let Alibaba, ByteDance and DeepSeek buy a capped number of Nvidia H200s, easing the shortage while holding limits on volume and use case. Both governments have now landed on the same instrument: allow access, cap it, keep a hand on the tap.

## Security tooling is the tell that something became infrastructure

On July 27, Nvidia launched the Open Secure AI Alliance with Microsoft, IBM, Red Hat, Cisco, Hugging Face, SK Telecom and more than thirty other organisations, focused on open models and agent harnesses, prompted by the sandbox breach involving OpenAI models and Hugging Face.

Nobody builds shared security tooling for a research artifact. They build it for a dependency. The vocabulary moved first: our own read of ninety days of coverage found "open weights" gaining share while "agentic" lost it.

Some hedging is warranted. A lobbying letter is not a law. The Chinese discussions were reported as discussions, not decisions. And Moonshot released K3 under a Kimi specific licence rather than a standard open one, which is its own quiet form of control.

## What to do with this

**Treat "open" as a policy variable with a date attached, not a property of a file.** China demonstrated in July that it will set a date and let products adjust around it: its interim measures on humanlike AI services took effect on July 15, and ByteDance and Alibaba pulled user configured agent personas ahead of the deadline. If a Chinese open weight model sits in your stack, your real exposure is regulatory and it sits upstream of the licence. Pin your version and keep your own copy of the weights you depend on.

**Watch one signal instead of the whole debate.** Does the next frontier release from Moonshot, Zhipu or Alibaba ship weights at all? That one decision will tell you more about where AI policy is heading than any letter to Washington. We are tracking it on the open weight policy trend.