OPEN-WEIGHT AI IS ENTERING A NEW ERA

Last Friday, 25 companies urged Washington to avoid 'premature restrictions' on open-weight AI models. The list of signatories has now passed 230.

Then on Monday, Moonshot openly released Kimi K3's secrets to all. Its weights - the trillions of numerical settings its training run established that define how the model works - can be downloaded by anyone.

The significance? Open weights have never been as close to the frontier as Moonshot’s Kimi K3. To the point that it’s making many uncomfortable.

But most organisations won’t download K3 and run it themselves. The infrastructure and expertise required remain considerable. And most workflows don’t require it.

But clouds, specialist operators, governments and well-resourced companies can now copy it and run it independently. It reduces the cost of near-frontier performance and increases how much control you can have over the served system. Which is direct and heavy competition for the frontier labs piling hundreds of billions into development.

So why is so much of corporate America mobilising behind open weights?

Partly money. A portable model needs chips, clouds, software and integrators. A glance at the original signatories shows most fit one of these camps.

Partly technology adoption. US leadership, the letter argues, depends less on who owns one frontier model than on whether AI reaches factories, hospitals and ordinary businesses. Greater choice across model sizes and deployment options can only help that.

Washington’s reaction, however, has been different. It worries about advanced capability that nobody can suspend. And while it could regulate US-based open-weight models, it cannot prevent foreign rivals from continuing to develop theirs.

So what is left, other than a ban on their use?

Well, the industry retorts, barring US companies from weights the rest of the world can download reduces nobody's risk. Frontier models in the hands of bad actors is the actual concern. And banning Chinese open weight models in America isn’t going to help that.

Yet it would also come at great cost. In complex regulated environments, you often need more control. And open weights make that control a choice.

You can choose the model and the operator. Then select where it operates - inside the plant, disconnected from the public internet or within the country where the data must remain.

Without access to open weight models, choice becomes constrained. Economic use cases are curtailed. And the rate of adoption across the economy dips.

But hey - Mr Trump would have greater control. And we would have less.