What are open weight AI models and why are they dividing Big Tech?

A cyberattack incident last week brought a technical AI term into the middle of a much bigger debate over who should control advanced artificial intelligence.

When a rogue OpenAI model accessed Hugging Face’s systems during a test, the company turned to GLM 5.2, an open-weight model developed by China’s Z.ai, to analyse and block the attack. Because Hugging Face could run the model on its own infrastructure, it did not have to send sensitive data, system information or login credentials to an outside provider.

The incident raised fresh questions about who controls advanced AI, who should have access to it and whether more openness makes AI safer or riskier.


Also read: Anthropic says Claude AI hacked three companies during cyber tests

What is an open weight model?

AI models learn by adjusting billions of numerical parameters during training. These parameters, called weights, help the model understand a prompt and decide what response to give.

In an open weight model, the developer makes these weights available to others. Anyone can download the model, run it on their own systems and adapt it for specific tasks.

This gives users more control than a closed model that can only be accessed through the developer’s service. Users can choose where its data is processed, customise the model and avoid depending on a single provider. Users can also use smaller, cheaper models for simple tasks instead of using an expensive frontier model every time.

However, an open weight model is not the same as an open source model.

Why open weights are not open source

The terms are often used interchangeably, but they are not the same.

The Open Source Initiative (OSI), a US-based nonprofit, says releasing weights alone does not reveal how a model was created. Developers may still not disclose the training code, training data, details about how the data was cleaned and checkpoints from the training process.

That makes it difficult for others to recreate the model or find security flaw came from. Under the OSI’s definition, open source AI should give users the freedom to use, study, modify and share the system, along with access to the material needed to make meaningful changes.

"Open Weights alone fall short of this because they do not provide the underlying training process, code, or comprehensive data details required for full-fledged use, study, modification, and sharing," the OCI said.

Big Tech and Anthropic divide over risk

An industry letter signed earlier this week by Microsoft, Amazon, Google, Meta, Nvidia, OpenAI and others argued that open weight models can widen access to AI, increase competition and reduce dependence on a small group of providers.

The signatories said startups, universities, businesses and public institutions should be able to use advanced models without having to train one from scratch. The letter also argued that greater access could improve safety by allowing more researchers to test models, identify weaknesses and develop safeguards.

Also read: OpenAI's Sam Altman discusses rogue agent and new AI models with US senators

Anthropic chief executive Dario Amodei has taken a more cautious position.

In a separate letter, Anthropic CEO Dario Amodei said his company does not support a complete ban on open-weight models and considers models without dangerous capabilities a public good.

However, he disagrees with the idea that broader access always benefits defenders more than attackers. Once powerful weights are released, they cannot be recalled, while guardrails and usage monitoring become harder. In cyberattacks or biological threats, attackers may gain more from broad access than defenders.

Amodei instead supports tighter chip controls, action against industrial-scale distillation and mandatory safety testing for sufficiently capable models, whether open or closed.