# Open-Source / Open-Weight AI Glossary

The canonical language for this workspace. Terms are promoted here only once they can be used correctly — this is a record of compressed understanding, not a dictionary to read cold. All lessons adhere to these definitions.

## Terms

**Open-weight model**:
A model whose trained parameters (weights) are published for download, so you can run it on your own hardware. Distinct from *open-source* in the strict sense — the training data and code may not be released.
_Avoid_: open-source model (use only when data + code are also open), free model

**Weights / parameters**:
The numbers learned during training that define a model's behaviour. "An 8B model" means ~8 billion of them. Running the model means loading these numbers into memory and doing arithmetic with them.
_Avoid_: the model's brain, variables

**Precision**:
How many bytes are used to store each parameter (e.g. FP16 = 2 bytes, 8-bit = 1 byte, 4-bit ≈ 0.5 byte). Lower precision = less memory, slightly less quality.
_Avoid_: accuracy, resolution

**Quantization**:
Compressing a model by storing its [[weights]] at lower [[precision]] than they were trained in, to fit larger models on smaller hardware. (Deep dive comes later — placed here as the term first appears in lesson 0001.)
_Avoid_: compression, shrinking

**VRAM**:
The memory on a GPU. The binding constraint for running a model: the weights (plus overhead) must fit in VRAM, or the model won't load / will be painfully slow.
_Avoid_: GPU storage, graphics memory