Anthropic reignites the open-model fight with a report on GLM-5.3

Anthropic published a report late last month arguing that GLM-5.3, released in August by the Chinese company Z.ai, poses a significant cyber-risk. The core claim: the powerful model is freely available for anyone to download and modify. It is the latest salvo in a long-running dispute over whether frontier models should be as open as GLM-5.3 or kept behind the doors of the big labs.

The debate plays out in public, and a lot of money is on the table. Nvidia chief executive Jensen Huang, who benefits when people buy hardware to run open models, enthusiastically pushes for their development in the United States. On the other side, Anthropic chief executive Dario Amodei, whose company could be hurt by cheap, powerful alternatives, advocates a more cautious approach. But beneath the commercial interests lies a fundamental disagreement. "I think a lot of where you land on the openness debate is shaped by where you land on the AI debate in general," says David Gray Widder, an assistant professor at the University of Texas at Austin's School of Information. Those who worry about AI safety warn that open models are easily exploited by bad actors; those focused on practical potential argue that such power should not sit behind the closed doors of a few large companies.

Start with the basics. A closed model — the kind that powers ChatGPT and Claude — is accessible only through a website, an app or an API. An open model, by contrast, can be downloaded directly: its internal architecture, encoded in a long string of numbers known as model weights, is public. These are sometimes called open-weight models to distinguish them from models that also release additional information such as training-process details.

For now, closed models lead on capabilities. As of publication, all eight top spots on Artificial Analysis's LLM leaderboard are held by closed models, although open models trail by only a few months of development. But there are good reasons to prefer an open model. Because they are downloadable, they can run on local machines or drives. They can also be retrained to behave differently, provided you have the expertise and resources.

If you are an academic researcher working to understand and improve AI systems, you almost certainly need open models. With closed models, researchers can mostly send prompts and observe responses. Open models let you peer into the internal mechanism, the chain of thought the model uses to reason through a problem before answering, combine multiple models into new systems, compress them for higher efficiency, and fine-tune them for new applications. Businesses have reasons to prefer open as well: even if you rely on an external provider to run them, open models typically cost less than closed ones.