Open weights
A model whose trained parameters are published for download, so you can run it offline and fine-tune it - narrower than ‘open source,’ which would also include the training data and recipe.
Open weights means the company that trained a model published the actual trained parameters - the billions of numbers that are the model - so you can download the file and run it on your own hardware instead of renting the model through someone else’s API. Qwen, Llama, Mistral, DeepSeek, and Gemma are open-weight; GPT-4 and Claude are not.
It is a narrower thing than “open source,” and the difference matters. Open source, in the traditional sense, means you get everything you’d need to rebuild the thing yourself: for a model that would be the training data, the training code, and the recipe. Open weights usually means you get only the finished numbers - the cake, not the ingredient list or the instructions. You can eat it, slice it, even re-frost it with a LoRA fine-tune, but you can’t rebake it from scratch, and you often can’t see what went into it. Many “open” models also ship with a license that restricts commercial use or bans certain applications, which is a further step away from what open source classically meant.
What you can do with open weights is most of what people actually want: run the model offline with no API bill and no data leaving your machine, inspect and measure its behavior directly, shrink it with quantization into a GGUF file that fits on a laptop, fine-tune it into a narrower or weirder version of itself, and keep running it unchanged for as long as you like - no vendor can deprecate a model you already have on disk. That last point is the quiet one: an open-weight model is the only kind nobody can take away from you.