Base, instruct & reasoning

Pretraining produces a base model; fine-tuning turns it into an instruct or a reasoning model. All three share one architecture and decode cost; only the behavior changes.

Base models#

A base model predicts the next token and nothing more: ask it a question and it may answer with three more questions. That makes it wrong for chat and right for raw completion, like filling the middle of a code file. The catalog ships instruct builds.

Instruct models#

An instruct model is a base model fine-tuned on instruction and response pairs, then aligned to human preferences. It reads the system, user, and assistant turns a chat is built from, follows directions, and handles tool use. A model listed with no qualifier is instruct.

Reasoning models#

A reasoning model writes a visible run of intermediate steps, then commits to a conclusion. The behavior comes from reinforcement learning against problems with checkable answers, so the thinking earns its keep on math, logic, and hard multi-step code. You pay in tokens and latency.

Reasoning, distilled#

DeepSeek’s R1 distills bake this behavior into an ordinary dense model: an R1-Distill-Qwen runs like a same-size Qwen and reasons like R1. Structured tool calls are not what they were tuned for.

Which to pick#

Default to instruct and step off it only for a reason; Choosing a model names the tuning behind each recommendation.

From the local API, the choice is just the model you name:

# instruct: the default for chat and tools
conifer serve --model qwen3-8b

# reasoning: step-by-step on a dense backbone
conifer serve --model deepseek-r1-distill-qwen-14b