AI glossary · Beginner · Language models
Base and Instruct models
A Base model has only completed pre-training and continues whatever text it is given, while an Instruct model has additional post-training that refines it to follow instructions and hold conversations.
About this term
- Level
- Beginner · Core ideas for first-time readers
- Category
- Infrastructure & using models
- Fields
- Language models
- Added
- 2026-10-12
How it works
A Base model has gained knowledge and language ability by learning to 'predict the next token' on vast amounts of text, but when asked a question it may continue with similar questions instead of answering. An Instruct model is a Base model that has been given SFT on instruction–response examples and refined to match preferences using methods such as RLHF and DPO. So when you ask it a question in the designated chat template, it answers conversationally. Most models behind today's chat services are Instruct-type models.
Base vs Instruct
Base modelPre-training only
- Continues text
- May answer a question with more questions
- Starting point for further training and research
Instruct modelPre-training + post-training
- Follows instructions and converses
- Ask in the chat template format
- Ready for use in services
공통 Knowledge and language ability come from pre-training
How to tell them apart by name
If a Hugging Face model name has a suffix such as Instruct, Chat or -it, it is usually a model refined to follow instructions. If it lists only the size with no suffix, or says Base, it is often a pre-trained model. However, naming conventions differ by company, so ultimately you need to check the model card. Whether a chat template is included is also a good clue.
Things to know
To plug a model straight into a service, choosing an Instruct model is the default. For extensive fine-tuning on your own data from the start, or for research, people sometimes choose a Base model. A Base model is not 'less smart'; it has the abilities but they are drawn out differently. Safety training is mainly added in the post-training stage, so Base models may be weaker at refusals and safe responses.
Example
A developer choosing an open-weight model for an internal chatbot finds two repositories of the same size on Hugging Face. One has Instruct at the end of its name and the other does not. Since it will be used for conversation right away, they pick the Instruct one. Meanwhile, as in research covered on this site, a result showing that simply fixing the starting token of the response brought a Base model's reasoning performance close to that of a model trained with reinforcement learning suggests the ability is already latent inside the Base model.