What is Base model?

A base model is the raw result of pretraining. It has absorbed patterns from an enormous body of text, code and other data, but it has not been taught to behave like a helpful assistant.

What it does

A base model predicts what comes next. Give it the start of a sentence, an email or a block of code and it continues in the most likely way. Ask it a question and it may answer, or it may produce more questions, because a list of questions is a plausible continuation. It is a powerful pattern engine with no sense that you are waiting for a reply.

How it becomes an assistant

Labs turn a base model into a chat product with post-training. Typical steps include fine-tuning on example conversations and reinforcement learning from human or automated feedback. That stage teaches the model to follow instructions, refuse harmful requests, format answers and, in a reasoning model, work through problems step by step. The result is often called an instruct or chat model.

Think of a base model as a well-read graduate with no job training. Post-training is the onboarding.

Why labs measure base models separately

Because pretraining is the most expensive stage, a base model is where a lab's investment in data and compute shows up. Researchers compare base models on prediction error and on benchmarks, without the effects of later tuning. That makes it easier to judge how efficiently a lab trained, though it is also something a lab can report in ways that are hard to check from outside, so look for independent results.

Why it matters to you

• Open releases. Some labs publish the base model alongside the tuned one. Builders often prefer the base model as a starting point for their own fine-tuning.
• Parameter counts. A model's headline size, such as the total parameters in a mixture of experts, usually describes the base model.
• Behavior. If a model rambles or ignores your question, you may be using a base model when you wanted an instruct version. Check the name on the download page.