With everyone discussing the world’s best AI models, it’s important to humanize the language we use to describe them. Understanding these key terms will help make conversations about large language models (LLMs) more accessible.
Model Family
Represents the brand of the model. There may be many models named similarly but they all belong to one "family". Like how Toyota has multiple cars. Llama has many types of models that belong to the Llama family.
Parameters
The more parameters a model has, the more detailed and complex it's understanding of it's training data. Thus having more detailed knowledge on individual topics which result in a more reliable and detailed answer.
Context Window
How much an AI can 'remember' at once while generating a response. A larger window means the LLM will have better memory when processing your next request... while taking context from previous requests into account.
Example:
Take the deepseek-r1 model, which has been generating a lot of buzz. Its model family is “Deepseek,” and the specific model is “R1,” with multiple versions available.
On Ollama model details page it shows the deepseek-r1 model as having the following parameter options: 1.5b, 7b, 8b, 14b, 32b, 70b, and 671b.
- The
1.5bbeing the least knowledgable. - The
671bmodel being the most knowledgable.
The 671b model boasts a 128k context window, meaning it can remember around fifty 1,000 to 2,000 word essays at once.
Hopefully, this helps distill AI terminology into simpler, more understandable concepts.