Definition
A base model is a large language model that has gone through only the first stage of training: pre-training on a vast corpus of text, learning to predict the next token. It has rich knowledge of language and the world, but it has not been taught to behave like a helpful assistant. Given a prompt, it simply continues the text in the most statistically likely way.
What a base model actually does
Because its sole objective is text completion, a base model’s behaviour can feel unintuitive. Ask a base model “What is the capital of France?” and it might answer “Paris” — or it might continue with “What is the capital of Germany? What is the capital of Spain?” because that pattern is also plausible in its training data. It is completing text, not following an instruction.
From base to instruct to chat
Modern assistants are built on base models through additional training stages:
- Pre-training produces the base model — broad knowledge, pure completion.
- Instruction tuning fine-tunes it on examples of instructions and good answers, producing an instruct model that follows commands.
- Alignment (often RLHF) further shapes tone, safety, and helpfulness, producing the chat model end users interact with.
Each stage builds on the previous one; the base model is the foundation that everything else is layered onto.
When base models are useful
Despite not being assistant-ready, base models have real uses. They excel at raw text completion and creative generation, and they serve as the starting point for custom fine-tuning — a team can take a base model and align it toward their own domain rather than fighting an existing assistant persona. Researchers also prefer base models when studying a model’s raw capabilities, uncontaminated by alignment.
Why it matters
The base model is the raw material of every chatbot you use. Understanding the distinction explains why providers ship separate base and instruct variants, why fine-tuning usually starts from a base checkpoint, and why an unaligned base model behaves so differently from the polished assistant most people expect.