Whether a model's trained parameters are public or locked behind an API changes who controls it, where it runs, and what you can do with it.
When a model finishes training, everything it learned is stored in its weights, the billions of numbers that make up its parameters. Whether those numbers are public is the dividing line between open and closed models.
Open-weight models, such as Meta's Llama, Mistral's releases, and DeepSeek, publish the weights for download. Closed models, such as OpenAI's GPT and Anthropic's Claude, keep them private and expose the model only through an API.
You can download the model, run it on your own hardware or a host of your choosing, inspect it, and fine-tune it. You are not dependent on one company staying online or keeping prices flat. The tradeoff is that running it well is your problem, and that takes real infrastructure.
You call an API and someone else handles the hardware, scaling, and updates. Frontier closed models are often the most capable available, and you get them with almost no setup. The tradeoff is less control: you cannot see inside, you cannot self-host, and your access and pricing depend on the provider.
Publishing weights is not the same as publishing everything. The training data and the exact training recipe are almost always kept private, and many "open" licenses restrict commercial use. You can run and adapt the model, but you usually cannot fully reproduce it from scratch.
With open weights you can run a model entirely on infrastructure you control, so sensitive data never leaves your environment. Cost becomes about hardware rather than per-message fees, which can favor either option depending on your volume. Closed models trade that control for convenience.
A closed model is a restaurant: you order, they cook, you never enter the kitchen. An open-weight model is a published recipe with the finished dish included: you can serve it as-is, tweak it, or cook it in your own kitchen, but you have to run the kitchen.
Berges AI runs a mix, choosing the right model per task rather than betting on one lab. Serving open-weight models is part of how it can offer strong options without locking you to a single provider. The model pages spell out which is which.
Try Berges AIThe gap has narrowed a lot. Top closed models often still lead at the frontier, but strong open-weight models are competitive for most everyday tasks, and you can run them yourself.
Smaller ones, yes. Large ones need serious GPUs or a hosting service. The bigger the model, the more memory and compute it takes to run at usable speed.
Not quite. Open weights means the parameters are downloadable. Truly open source would also include the training data and code, which is rare. Read the license, since many open-weight models restrict how you can use them.