You're viewing revision R0 from Nov 17, 2026. See the current version.

Most people meet AI through a chat box. You type, it answers, and the company behind it decides what it will and will not say. That is a closed model. OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini are closed. You reach them through an app or an API. You cannot download the model and run it yourself.

Other models are open weight. Anyone can download the weights and run the model on their own hardware. Meta's Llama, Mistral, and DeepSeek are examples. The weights are the trained parameters, the numbers that shape the model's behavior. Having them means you can run the model without asking a company for access.

Open has layers

Open weight is not the same as open source. The Open Source Initiative says publishing weights alone does not meet its definition of open source AI. A true open source AI system also gives users the freedom to use, study, modify, and share it, with access to the training code and enough information about the training data to understand and rebuild the system. Many models called open source are really just open weight.

A third layer is openness about rules. Some open weight models still carry usage policies that restrict what you may do with them. A company can also be open about its refusals, publishing clear policies on what its models will not help with. That kind of openness is separate from whether the weights are downloadable.

More open is not automatically better

An open model is not safer or smarter by default. It is more transparent and more in your control. You can inspect it, run it locally, and keep your data on your own machine. You can also modify it. But the same freedom applies to everyone, including people with bad intentions. Closed models keep some decisions in the company's hands. Open models put them in yours. That is a trade, not a guarantee.

Sources: Wikipedia on open source artificial intelligence, Open Source Initiative, Wikipedia on large language models.

Inspired by Happy, written by Claude.

Revision history (R0 current)
RevDateWhat changedPage
R0Nov 17, 2026First published.view