Learn / Glossary
Glossary
The technical words that appear on the site, explained in one sentence.
22 terms
A
- Adoption
- How much it's really used: monthly downloads, popularity on OpenRouter or likes on Hugging Face.
- API
- A door for programs to use AI directly, paying for what they consume. It's what developers use.
B
- Benchmark
- A standard test to measure and compare models independently.
C
- Chatbot
- The app or website where you talk to an AI model, like ChatGPT, Claude, Gemini or Le Chat.
- Context
- How much text the AI can take into account at once (the conversation, attached documents…).
- Quantization
- Compressing a model (for example to 4 bits) so it takes up less memory, losing a little quality.
E
- Elo
- A score that comes from thousands of blind votes between two results, as in chess: higher = people prefer it more.
- Embeddings
- Turning texts or images into lists of numbers to search by meaning. Used for search engines and recommendations.
F
- Efficient frontier
- The models that nothing beats on both quality and price at once: they're the best options for each budget.
G
- GPU
- The graphics card. It's the component that runs AI models on a computer.
H
- Hugging Face
- The website where most open models are published and downloaded.
L
- License
- The terms of use of an open model: some allow everything (Apache, MIT) and others forbid commercial use.
M
- Maturity
- Whether a model is already proven: how many independent evaluations have measured it and how long it has been available.
- Model
- The «brain» of an AI: the file with everything it has learned. Each app (ChatGPT, Claude…) uses one or more models.
Or
- Open source
- A model whose «weights» can be downloaded: you can use it on your own computer or server without paying anyone (respecting its license).
- OpenRouter
- A service that gives API access to hundreds of models; it publishes rankings of the most used ones.
P
- Parameters
- The model's internal «connections». More parameters usually means more capability, but also more memory and cost.
- Weights
- The file with everything the model has learned. If they're «open», anyone can download them.
R
- Reranker
- A model that reorders a list of results to put the most relevant ones first.
S
- Self-host
- Running the model on your own machine or server instead of using another company's service.
T
- Token
- The piece of text an AI works with (roughly ¾ of a word). APIs charge per million tokens.
V
- VRAM
- The memory of the graphics card (GPU). It decides which models you can run on your computer: the bigger the model, the more VRAM it needs.
We couldn't find that term.