ExplainerIntermediate6 min read
What 'open' means when a lab says open weights
Open weights, open source, open research — three different things that all sound like the same thing. How to tell them apart in one read.
Why the words blur together
A lab publishes a model. The post says open. Within a day, half the coverage calls it open source and the other half calls it a leak of weights, and both sides are sure they are right. They are usually describing different parts of the same release.
The distinction is not pedantry. It decides whether you can audit the model, whether you can use it commercially, and whether you can be cut off from it later. The rows below are the ones that matter.
The words that do the work
- Open weights
- The model's parameters are published as a file you can download and run yourself. Nothing about the training data, the code, or the method is included. This is what most 'open' AI releases actually are.
- Open source
- A specific legal standard, set by the Open Source Initiative: the source, the license, and the build process, free of restrictions on use or redistribution. By that standard almost no frontier model qualifies, because the weights without the training recipe are not source.
- Permissive license
- A license like MIT or Apache 2.0: do what you want, including commercial use, just keep the notice. Model licenses modeled on these are the friendliest to build on, and the ones whose fine print is worth reading once.
- Copyleft
- A license like GPL: you may use and modify it, but anything you ship that includes it must share its own source too. Rare in model weights, common in the tooling around them.
- Base vs instruct
- A base model is the raw trained predictor — powerful, and quite happy to finish your prompt with something nobody asked for. An instruct or chat version has been fine-tuned to answer as asked. Downloading the base when you wanted the assistant is the most common first mistake.
- Quantization
- Shrinking the model's numbers from 16-bit to 8- or 4-bit so it fits on smaller hardware. It costs a little quality and a lot of RAM, which is usually the trade you actually want.
- GGUF and friends
- File formats for running quantized models locally. When a release ships in several of these, it is aimed at people running it on their own machines, which is a real signal of intent.
- Model card
- The document that ships with the weights: what was trained on, what it is for, and what it is not for. The honesty of a release lives here, and the good ones are genuinely short.