An open-weights model is an AI model whose trained numbers, the weights, are published for anyone to download and run. You do not have to call someone else’s API. You can host the model on your own servers and keep full control over how it behaves and where your data goes.
A helpful comparison is a recipe versus a finished meal. A closed model like the one behind ChatGPT is a meal delivered to your door: convenient, but you never see the kitchen. An open-weights model is the cooked dish handed over with permission to reheat, season and serve it however you like. Llama from Meta and Mistral’s models are well-known examples. Because you hold the actual AI model, you can run inference on your own machines and even fine-tune it on your own data.
It is worth being precise: open-weights is not the same as fully open-source. The weights are shared, but the original training data and full pipeline usually stay private. So you get the engine, not the factory that built it. The licence still sets limits too. Some open-weights releases restrict commercial use above a certain user count, so reading the terms matters before you build on one.
The honest catch is the running cost. A capable model needs a serious GPU, and keeping it patched and fast is real engineering work, not a one-time download. For a low-volume task, a paid API can still be cheaper than the hardware.
At TopDevs we reach for open-weights models when a client needs sensitive data to stay in-house or wants to avoid being locked into a single API vendor.
