Base, Instruct and Reasoning Models: What Each Is For

When you browse the Hugging Face Hub you will see the same model family released under several names, such as Llama-3.1-8B and Llama-3.1-8B-Instruct, or gemma-2-9b and gemma-2-9b-it. Hosted providers also list separate reasoning or thinking models. These are not interchangeable. They come from different training stages and behave differently. Picking the wrong one is a common and expensive mistake.

The mental model is a pipeline of training stages:

code

pretraining on huge text corpus -> BASE model

    -> supervised fine-tuning on instructions + preference tuning -> INSTRUCT model

        -> reinforcement learning on verifiable tasks (math, code, logic) -> REASONING model

1. Base models: next-token predictors

A base model (also called pretrained or foundation model) has only been trained to predict the next token on large amounts of text. It has broad knowledge and language skill, but it has not been taught to follow instructions or hold a conversation.

Ask a base model 'What is the capital of France?' and it may answer. It may just as easily continue with more quiz questions, because on the web a question is often followed by more questions. It is completing a document, not answering you.

What base models are for:

Starting point for fine-tuning. If you are training your own instruction-following or domain model, starting from the base avoids inheriting another team's chat style and refusal behaviour.
Pure text completion. Autocomplete, continuing a document in a fixed style, or generating synthetic data from a few-shot pattern.
Research and evaluation of what the model learned in pretraining.

How to spot them on the Hub: names without a suffix, or with -Base. The model card usually says 'pretrained' and has no chat template.

2. Instruct models: the default for applications

An instruct model (also called chat model, with suffixes like -Instruct, -Chat or -it) starts from a base model and is further trained with:

Supervised fine-tuning (SFT) on examples of instructions paired with good responses.
Preference tuning such as RLHF or DPO, which teaches the model to prefer helpful, harmless and well-formatted answers.

The result follows instructions, respects system prompts, uses a with roles, and supports features like tool calling and structured output.

Trade-off	Why it matters
More tokens	Thinking tokens are generated output and are usually billed, even if they are hidden from you. A short answer can cost thousands of tokens.
Higher latency	Time to the final answer can be seconds to minutes. This is a poor fit for real-time chat or autocomplete.
Less control	Some providers restrict or ignore temperature and top_p for reasoning modes.
Overthinking simple tasks	For classification, short extraction or simple lookups, they add cost and latency with little or no quality gain.

Task	Good default	Reason
Support chatbot, RAG answers	Instruct	Follows instructions and formats cleanly at low latency
High-volume classification or extraction	Instruct, often a small one	Cheap and fast. Reasoning adds cost without much gain.
Complex code generation or debugging	Reasoning, or hybrid with thinking on	Multi-step correctness matters more than speed
Agent planning over many tools	Reasoning for the planner, instruct for simple steps	Spend thinking only where decisions are hard
Fine-tuning your own domain model	Base, or instruct if you only have a small dataset	Base gives a clean start. Instruct already has chat skills.
Autocomplete or document continuation	Base or a code-completion model	Completion is exactly what base models do

Base, Instruct and Reasoning Models: What Each Is For

1. Base models: next-token predictors

2. Instruct models: the default for applications

3. Reasoning models: think first, then answer

Choosing between them

Checklist you should be able to explain