Fine-tuning takes a pre-trained language model and continues its training on a smaller, curated dataset. The model's weights are updated further through additional training steps to change its behavior, style, or domain competency — while retaining the broad language understanding from pre-training.
Trains from scratch on hundreds of billions of tokens. Teaches grammar, facts, and reasoning from the ground up. Costs tens to hundreds of millions of dollars. Done once by model providers.
Starts from a pre-trained model. Trains on thousands to hundreds of thousands of curated examples. Far cheaper — possible on consumer hardware with the right techniques. Done by developers to specialize behavior.
Output format — The model reliably responds in JSON, markdown, or a specific structure without needing to be instructed every time.
Domain terminology — Medical, legal, and financial jargon used naturally and correctly without prompting.
Tone and style — A consistent brand voice, communication style, or conversational register across all responses.
Reasoning patterns — Structured clinical reasoning, legal argument construction, or domain-specific problem-solving frameworks.
If you need the model to know something, use RAG.
If you need the model to act in a specific way, fine-tune.
Fine-tuning is a fragile mechanism for injecting facts — they can't be updated without retraining. It excels at changing behavior, style, and format patterns.
Ask the AI assistant about what fine-tuning is, how it differs from pre-training, or when to use it vs RAG.