Weights
also called Parameters
The numbers inside the model that training moved.
The learned numerical values on the connections inside a neural network. Modern models have billions; their count is the headline number in a model’s name.
People say "parameters" for the count and "weights" for the specific values. They are the same thing — the model’s entire knowledge is stored in them, in a form nobody can read directly.
Weights are the artefact that makes a model a thing you can possess. That is what makes open weights strategically different from an API: you can inspect them, fine-tune them, and keep running them after the vendor moves on. It is also why the adapters around them are where most of the practical differentiation now lives.
What you actually get with the download
A legal team asks what "open weights" means for a compliance review.
You get the trained numbers and the ability to run inference locally. You do not automatically get the training data, the recipe, or unrestricted commercial rights — those come from the licence, which varies by model and is worth reading before the architecture discussion happens.
- paperLoRA: low-rank adaptation of large language models ↗
Freeze the pretrained weights, train small low-rank matrices beside them. The origin of parameter-efficient fine-tuning, and why adapters are cheap to ship and swap.
- paperLlama 2: open foundation and fine-tuned chat models ↗
A released family from 7B to 70B plus chat variants, with the fine-tuning and safety work documented. A primary example of published weights — not a ruling on what "open source" should mean.
Parameters
How big the model is, counted in learned numbers.
Training
The expensive one-off process that sets the model’s parameters.
Quantization
Storing weights at lower precision to make a model cheaper to run.
Interpretability · Explainability
Working out why the model did that.