Instruct model
Also known as: instruction-tuned model, chat model, assistant model
A base model trained purely on next-token prediction doesn't know what to do with a question or instruction. If you prompt it with 'Explain how photosynthesis works,' it might just continue the sentence in a random direction. An instruct model has been fine-tuned on thousands of prompt-response examples and alignment techniques so it responds helpfully when given instructions.
The naming convention varies by lab. OpenAI uses 'GPT-4o' versus 'GPT-4-base.' Meta uses 'Llama-3-Instruct' versus 'Llama-3.' Anthropic's publicly released models are all instruct models by default. When you see an open-weight model on Hugging Face (an open-source model repository), you'll often find both a base checkpoint and one or more instruct variants.
For builders doing fine-tuning, the choice between starting from a base model versus an instruct model matters. Starting from an instruct model is usually faster and safer for products where you want conversational behavior. Starting from a base model gives more flexibility for domain-specific pre-training, but requires more work to make the result usable. Most practical fine-tuning projects today start from an instruct model.