Inspect the license, the actual deployment, and the responsibilities at each layer.
More room to examine and adapt
Access to model weights can support independent inspection, research, and local deployment. The exact permissions depend on the release and its license; “open weights” does not describe every legal or technical condition.
The deployment still needs safeguards
The application determines which tools, data, users, and external systems the model can reach. Adaptation can change behavior, so a base model’s evaluation should not be treated as a guarantee for a modified system.
Tradeoffs deserve a public conversation
Openness can support accountability and competition, while also enabling misuse. The useful question is which controls fit a particular capability and context, and who bears their costs.
Ask which parts are open
The Open Source Initiative’s AI definition covers freedoms to use, study, modify and share, with access to components needed to exercise them. A release may expose weights while imposing other restrictions or omitting information needed to reproduce training. Describe the actual access and terms instead of treating ‘open’ as a single binary feature.
Source: Open Source Initiative · Open Source AI Definition · 1.0Control brings operating work
Running weights yourself can allow you to choose hosting, access rules and adaptation. It also requires operating the system: patching dependencies, managing secrets, observing failures and handling user requests. Our interpretation is that control is valuable only when someone has the capacity and accountability to exercise it well.
Inspect a real model card
Meta’s Llama 3.1 card provides a concrete example of documented intended use, evaluations and release terms. Use a specific card to practice reading those details, while keeping separate the original release and whatever service, fine-tune or tool integration you actually encounter.
Source: Meta · Llama 3.1 model cardA situation to think through
A small organization considers hosting its own document assistant. Keeping inference on its infrastructure may suit its data policy, but the team must still decide who can query which documents and how logs are protected. The weights do not answer those operational questions.
Questions to take with you
- Identify the exact weights, license and modifications.
- Determine who can maintain the deployment and respond to incidents.
- Compare practical access and accountability, not only the open/closed label.
For more reading
The sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.
- Open Source AI Definition · 1.0
One explicit definition of openness, covering freedoms and access to more than model weights.
- Llama 3.1 model card
A concrete open-weight release: intended use, deployment considerations, evaluations and license terms.
- AI Agent Security Cheat Sheet
Practical guidance on tool permissions, memory isolation, oversight and agent failure handling.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence