Open Weights Are Becoming Less Open As AI Labs Add Conditions

Open-weight AI used to sound simple. Either the model weights were available, or they were not. That world is disappearing quickly.

Z.ai’s GLM-5.3 rollout is a good example of the new middle ground. The Chinese AI company has released GLM-5.3-Flash weights publicly on Hugging Face, but The New Stack notes that the model comes with a commercial catch: companies with more than $10 billion in annual revenue must pass a Z.ai security review before using the model or derivative works commercially.

That means the model is open enough for developers, researchers and many smaller companies to download and experiment with, but not open in the old unrestricted sense. A hyperscaler, large AI lab or global enterprise cannot simply treat it like a permissive Apache-style release. It has to deal with the license conditions first.

This is where the language around open AI is becoming messy. Open source traditionally means more than downloadable files. It involves rights to use, study, modify, redistribute and build on software without selective commercial gatekeeping. Open-weight models are different. They may expose the trained parameters while still restricting data, training code, commercial use, safety review or deployment rights.

Z.ai is not alone in moving toward this conditional model. Meta’s Llama releases have also mixed wide availability with licensing limits for very large companies. The trend suggests that AI labs want the distribution and ecosystem benefits of openness without giving the largest rivals a completely free commercial advantage.

From a business perspective, the logic is understandable. Training frontier or near-frontier models is expensive. If a lab gives away everything with no conditions, cloud giants and well-funded rivals may capture much of the value. Conditional open weights let the lab build developer goodwill while still keeping leverage over the biggest commercial users.

From a public-interest perspective, the tradeoff is more complicated. Open weights can make AI more accessible, especially for countries, startups, universities and developers that cannot afford closed frontier APIs. They also allow independent testing, local customization and more transparency than fully closed models. But restrictive licenses can limit who can deploy them at scale.

The security argument is also real. Z.ai had previously delayed wider release of GLM-5.3 because of concerns about cyber capabilities. Axios reported earlier this month that Chinese open-weight models were closing in on U.S. frontier systems for finding and exploiting security flaws. If a model can materially help attackers, a lab may argue that some commercial review is responsible rather than anti-open.

The danger is that every company will define openness in the way that best suits its strategy. A model may be marketed as open because developers can download it, even if the license blocks major commercial use or imposes opaque approval requirements. Readers and developers need to ask the next question: open for whom, under what terms, and at what scale?

This matters for Africa and other emerging markets. Open-weight AI can reduce dependence on foreign APIs and make local language, education, healthcare, finance and government tools more realistic. But if the best models become open only below certain revenue thresholds or under geopolitical review, access may still be shaped by corporate and national power.

The more honest vocabulary may be needed soon. Some models are open source. Some are open weight. Some are research open. Some are commercially restricted. Some are open only until the user becomes large enough to matter. Treating all of them as the same thing helps marketing departments more than it helps developers.

The future of open AI will probably be conditional, layered and political. That does not make it useless. It means the word open now needs careful reading. In AI, the license may matter almost as much as the benchmark.