Base Model vs Instruction-Tuned Model

Simple Definition

A base model is an AI model at the end of pre-training. It has learned to predict the next piece of text from enormous amounts of data, but it has not been taught to follow requests. An instruction-tuned model is a base model that went through further training so that it answers questions, follows instructions, and behaves like an assistant.

What Happens Between the Two

Pre-training produces the base model. The model reads trillions of tokens of text, code, and sometimes images, and learns language, facts, and patterns by predicting what comes next.

Post-training turns it into something you can talk to. It usually includes:

  • Instruction tuning: supervised training on examples of requests paired with good responses.
  • Preference training, such as RLHF, where the model learns which of several answers people prefer.
  • Safety and alignment training, including when to decline.
  • Tool-use training, so the model can call functions and work as an agent.
  • A chat template, the special formatting that marks which text came from the user and which from the assistant.

The difference shows quickly. Type “Write a haiku about rain” into a base model and it may continue as if reading a textbook, adding more writing prompts or an exercise list. An instruction-tuned model writes the haiku.

Example

On November 5, 2026, Baidu published ERNIE 5.5 Flash Base, a 280-billion-parameter open-weight model. Its model card describes it as the checkpoint at the end of pre-training, before any further training, and serves it through a completions endpoint rather than a chat one. Baidu positions it for domain adaptation, instruction tuning, and research. That makes it valuable to a team that wants to build its own specialized model from a strong starting point. Someone who downloads it expecting a ready-made assistant will be disappointed: it has not been taught to follow requests or to decline harmful ones. More in Frontier Open Models Explained.

Comparing Benchmarks

Base and instruction-tuned models are tested differently. Base models are usually scored with few-shot prompts, where examples of the task are placed in the prompt, on knowledge, reasoning, and coding tests. Instruction-tuned models are scored with chat formatting, often zero-shot, and on agent tasks with tools. ERNIE 5.5 Flash Base’s benchmark table, for example, compares it only with other base models. Setting a base model’s score next to a chat model’s score from a different setup tells you very little.

Which One Do You Want?

  • To use a model directly, in an app, a chat, or an agent: an instruction-tuned model.
  • To build your own assistant or domain model, or for research: a base model gives you a clean start without another lab’s assistant behavior baked in. You will need training data, compute, and your own safety work.
  • To adapt a model to a narrow task: often an instruction-tuned model with light fine-tuning is cheaper and gets there faster.

Many open-weight releases publish both, usually marked with names like “Base” and “Instruct.” Read the model card before you download.

Continue learning

Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.

See AI terms in action

Browse practical AI workflows that use the concepts in this glossary.

Last updated: