Base, Instruct, Reasoning, Hybrid: the LLM model types, explained
Not all LLMs are the same kind of thing. Knowing the difference between a base model, an instruct model, a reasoning model, and a hybrid tells you which one to reach for.
"Which model should we use?" usually gets debated as a brand question — Claude vs GPT vs Llama. But there's a more useful axis underneath: what type of model it is. Base, instruct, reasoning, and hybrid are different tools for different jobs.
Base models
A base model is the raw result of pretraining on a huge corpus. It does one thing: predict the next token. It'll happily continue your text, but it hasn't been taught to follow instructions or hold a conversation — ask it a question and it might just write more questions. Base models are the foundation everything else is built on, and mostly of interest if you're doing your own fine-tuning.
Instruct / chat models
Take a base model, fine-tune it on instructions, and align it to human preferences (typically with RLHF), and you get an instruct or chat model. This is what you actually use day to day — it follows directions, stays on task, and converses. When someone says "GPT-4o" or "Claude," they mean this.
Reasoning / thinking models
Reasoning models are trained to think before they answer — producing an internal chain of reasoning, often via reinforcement learning, before giving a final response. They're markedly better at hard, multi-step problems: maths, code, complex planning. The trade-off is more time and more tokens, so they're overkill for simple tasks.
Hybrid models
The newest models blur the line: they can switch thinking on for a hard problem and off for a quick one, giving you fast responses and deep reasoning from a single model. It's the direction the field is heading — one model that dials its effort to the task.
- Base — for fine-tuning and research, not direct use.
- Instruct / chat — the default for almost everything.
- Reasoning — for genuinely hard problems worth the extra latency and cost.
- Hybrid — one model that adapts its effort to the task.
Choosing a model isn't only picking a brand. It's picking the right type — and paying for reasoning only when the problem deserves it.