A base model, an instruct model, and a reasoning model often belong to the same model family. The labels indicate how a checkpoint was produced, how it is meant to be used, and what kind of behavior we should expect from it.

Base model

A base model is the checkpoint obtained from broad next-token pretraining. It can complete text and contains the general capabilities learned from the training corpus. However, a raw base model may continue the wording of a prompt instead of treating it as a request from a user. This makes base checkpoints useful for research and further finetuning, but usually inconvenient for a chat application.

Instruct model

An instruct model usually starts from the base checkpoint and receives additional training on instructions and desired responses. Many training pipelines also use preference data after this instruction-finetuning stage. The resulting model is more likely to answer a question directly, follow a requested format, and use the chat template expected by the model family.

Reasoning model

A reasoning model is a less standardized category. In current model families, it usually refers to a checkpoint or inference mode shaped for problems that require several intermediate steps. The post-training recipe may involve reinforcement learning on verifiable tasks, distillation from another reasoning model, or both. At inference time, these models often use a larger token budget before returning the final answer.

That extra budget can help with math, programming, and other multi-step tasks. It also increases latency and token cost, and it may add unnecessary work for a simple request.

The Qwen overview in the repo shows how a modern family can expose multiple behavioral variants rather than just one monolithic model type

The Qwen material in the repo provides a concrete example. Some releases offer separate base and instruct checkpoints, while others expose a thinking mode through the chat template or generation settings. The exact packaging changes between releases even though the broad distinction remains useful.

The repo’s Qwen materials also include variants such as coder and flash-style models, which reinforces that modern model families are increasingly packaged around distinct use cases and response behaviors

Names such as Coder or Flash describe another axis, namely the intended use or deployment tradeoff. A coder model can still be a base, instruct, or reasoning-oriented checkpoint. Model-family labels therefore don’t form a strict three-box taxonomy.

For choosing a checkpoint, I use the following rule of thumb:

  • choose a base model for custom finetuning, representation studies, or raw text continuation
  • choose an instruct model for ordinary chat and prompt-following tasks
  • choose a reasoning model when a difficult task benefits from a larger inference budget