A base model is the raw pretrained text-completion model; an instruct (or chat) model has been further tuned to follow instructions and hold a conversation.
One of the most common early stumbles when downloading open models is grabbing the wrong variant and getting bizarre results. Most model families ship in (at least) two forms — a base version and an instruct/chat version — and the difference is fundamental to how they behave. Picking the right one is a two-second check that saves a lot of confusion.
A base model (also called a foundation or pretrained model) is trained on one enormous objective: predict the next token across vast amounts of text. That's it. This gives it broad knowledge and strong language ability, but it only knows how to continue text — it has never been taught to be helpful, to answer questions, or to follow instructions. So if you ask a base model "What is the capital of France?", it might answer correctly, or it might continue as if it were reading a quiz and generate more questions, or drift into unrelated text. It's completing a document, not talking to you. Base models are the raw starting point: they're what you use for further training, for certain pure text-completion tasks, and as the foundation to fine-tune.
An instruct or chat model takes that base and fine-tunes it to behave like an assistant. This second stage typically involves instruction tuning (training on many examples of instructions and good responses) and often preference tuning such as RLHF (reinforcement learning from human feedback) or DPO, which align the model toward responses people prefer. The result follows directions, answers questions directly, stays on task, and respects a chat format with distinct roles (system, user, assistant). This is the version almost everyone actually wants — any assistant-style, question-answering, or conversational use calls for the instruct/chat model.
The practical way to tell them apart is the name. Look for a suffix like "-Instruct," "-Chat," "-it" (Google's Gemma uses "-it" for instruction-tuned), or similar — that's the ready-to-talk version. The plain name with no such suffix, or one labeled "-base" or "-pt" (pretrained), is the base model, which expects you to either fine-tune it or use it for raw completions with careful prompting. Downloading a base model when you wanted a chat model is a classic beginner surprise: it seems "broken" or "won't follow instructions," when in fact it's working exactly as designed — you just have the wrong variant.
There's also a chat-formatting detail worth knowing: instruct models are trained with a specific chat template (special tokens that mark where the system prompt, user turn, and assistant turn begin and end). Friendly local tools like Ollama and LM Studio apply the correct template for you automatically, which is another reason they're easy to start with. If you use a raw engine directly, you may need to format the prompt with the model's expected template, or it will underperform.
There are a couple of intermediate cases worth knowing. Some models ship additional variants beyond the basic base/instruct pair — for example a "reasoning" variant tuned to think step by step, or a coder variant fine-tuned on source code. These are still, at heart, fine-tuned descendants of a base model, specialized for a particular use. And a few base models are surprisingly usable with careful few-shot prompting (showing examples in the prompt), but for anything conversational the instruct variant is far less fiddly and far more reliable.
Why do publishers release the base model at all, if most people want the instruct version? Because the base is the more flexible foundation for building. Researchers and developers who want to fine-tune a model for a specialized behavior often prefer to start from the base, since it hasn't already been shaped toward general chat behavior that they'd have to work around. The base is the raw clay; the instruct model is a finished, general-purpose tool made from it. If you're just using a model, take the finished tool; if you're building a custom one, you might want the clay.
How does this connect to the rest? Instruct models are the product of fine-tuning applied to a base model, and if you want to specialize a model for your own use, you'll choose which variant to start from depending on your goal — usually fine-tuning an instruct model for assistant tasks, or a base model when you're doing heavier customization. LoRA adapters likewise target one or the other. When you browse Spanvero's catalog at /models/, the assistant-ready instruct/chat variants are the ones most people should run; the honest cost to run any of them — locally, on a rented GPU, or via your own API key — is on each model's page and in the calculator at /calculator/.
Fine-tuning · LoRA · Inference · Parameters (the "B" / billions) · Ollama · Quantization
All explainers → · Browse models →
Open the free Spanvero advisor → · Honest, $0-markup. © 2026 Cynosure LLC.
A short email of real AI price moves, straight from the daily log — no hype. We're collecting the list now; the first issue goes out when it opens. Unsubscribe with one click.