Learn
Glossary
The same words as the Words lesson, in a list you can jump back to.
- Parameters
- The size of the model, in billions. A mixture-of-experts model lists the total you store and the active count used per token. Memory follows the total.
- Quantization
- A smaller, slightly lossy copy of the weights. Q4_K_M is the default on the memory table. Q8_0 is larger and closer to the original weights.
- GGUF
- The file format llama.cpp and many local runners use. It is not a model family.
- Context length
- How much of the conversation the model can keep. The memory table here is computed at 8,192 tokens unless a measured row uses another length.
- Tokens per second
- How fast the model writes. A measured figure names the machine, the quant, the context, and the date. Otherwise the cell says not yet measured.
- Computed
- Memory math from the published parameter count and attention config. Not a measurement of the file on disk.
- Vendor-reported
- A figure from the maker’s card, kept separate from a measured run and from the computed table.
- Sold and asking
- A sold price is what a listing closed for. An asking price is what a seller wanted. They are dated, and they are never added together.