Weights, Parameters and Model Families

Decoding a model name

Open-weight models are published by labs in families (Llama, Qwen, Gemma, Mistral, Phi and others), each in several sizes measured in parameters: 1B, 3B, 7B, 8B, 14B, 70B and so on. More parameters generally means better knowledge and reasoning but more memory and slower generation. Most families ship two flavours. A base model was only pre-trained to continue text; ask it a question and it may continue with more questions. An instruct (or chat) model was further fine-tuned to follow instructions in a conversational format, which is what you want for assistants. Some names also carry the context length or a specialty such as coder or vision. Finally, check the licence: open weights does not always mean free for every commercial use, and each family has its own terms.

Engine size on a bike

A 100cc bike and a 350cc bike from the same company share a design, but the bigger engine pulls harder and drinks more fuel. Model sizes within a family work the same way: same architecture, more parameters, more capability, more memory and power.

Base vs instruct trap

If a model rambles, repeats your prompt or ignores instructions, check that you pulled the instruct variant and that the correct chat template is applied. A base model behind a chat UI is a common beginner mistake, and it looks like the model is "dumb" when it is simply the wrong flavour.

त्वरित जाँच: You want a local assistant that answers questions in a chat. Which variant should you choose?

• The base model, because it is the original
• Any variant; they behave identically
• The largest model regardless of RAM
• The instruct (chat) fine-tuned variant