Open Weights Model

An open weights model is one where a company releases the entire set of values an AI ended up with after training, putting them online so anyone can download the model and run it on their own computer or server. Those 'values' are the huge collection of numbers stored inside the AI that determine how it makes judgments.

A restaurant analogy makes this easier to grasp. A closed model is like a restaurant where customers can only order and eat a dish there. The recipe is never shared, and you pay every time you eat. An open weights model is like publishing that recipe in full. Anyone can take it and cook the same dish in their own kitchen — though if the kitchen (computing power) isn't good enough, cooking it can be slow or difficult.

How companies feel about this approach depends on their business model. Companies that make money selling kitchen tools have little to lose from releasing the recipe, so they tend to welcome it. Companies that make money by serving the dish themselves, on the other hand, tend to insist on keeping things closed, since revenue takes a hit the moment the recipe gets out.

Article 1 contrasts Meta, Google, and NVIDIA, which actively back open source and open weights models, with OpenAI and Anthropic, which stick to a closed approach. Article 2 introduces a conversation in which vLLM's lead maintainer shares experience running open weights models in production on large-scale GPU clusters. A common misconception is treating 'open weights' and 'open source' as the same thing — an open weights model usually only releases the internal values, not the training data or the full training methodology.

To check whether a model is open weights, look for phrases like 'open weights' or 'open source license' in the official announcement. Then check whether you can download and run the model yourself on a personal computer or server, or whether it's only accessible through a specific company's service — that distinction tells you whether it's truly open weights.