What a chat model is
A chat model is a large language model that has been specifically trained to hold a conversation. Instead of just continuing a block of text, it interprets input as a sequence of messages with roles — typically system, user and assistant — and produces a helpful assistant reply that respects the conversation so far. Products like ChatGPT, Claude and Gemini are all powered by chat models. Without this conversational training, a raw model is far harder to use directly.
Base models vs chat models
Under the hood, every chat model starts as a base model — a network trained on huge amounts of text to predict the next token. A base model is a pure text predictor: if you type a question, it might continue with more questions rather than answering, because completion is all it learned to do.
A chat model is a base model that has gone through additional training to:
- Recognise the conversational message structure.
- Follow instructions instead of merely continuing text.
- Answer helpfully, stay on topic, and refuse unsafe requests.
This is why you can simply ask a chat model a question and get a sensible reply, while a base model needs careful prompting to coax the same behaviour out.
The chat format and templates
Chat models expect messages tagged with roles:
- system — background instructions and persona (the system prompt).
- user — what the human says on each turn.
- assistant — what the model replies on each turn.
Before this reaches the model, a chat template wraps each message with the model’s specific special tokens and formatting. Each model family has its own template, and using the correct one matters — feeding a model the wrong chat format can measurably hurt its responses. APIs usually apply the right template for you, but it becomes important when running open models yourself.
How chat models are trained
Turning a base model into a chat model generally involves two stages on top of pretraining:
- Supervised fine-tuning (SFT) — the model is trained on many example conversations written or curated by humans, learning the input/output shape of a good assistant turn. This is also called instruction tuning.
- Reinforcement learning from human feedback (RLHF) — humans rank competing responses, a reward model learns those preferences, and the model is fine-tuned to produce answers people prefer. This pushes it toward being more helpful, honest and harmless.
Some labs use related techniques (such as direct preference optimisation) for the second stage, but the goal is the same: align the model’s behaviour with what users actually want from a conversation.
Why it matters
The chat-model pipeline is what makes today’s AI assistants feel cooperative rather than like an unpredictable autocomplete. Understanding the system/user/assistant structure also helps you prompt better — you control the system role to set rules, the user role to ask, and you can even supply prior assistant turns as in-context examples to steer the style of what comes next.