# Modelfile Basics: FROM, SYSTEM and PARAMETER — Local LLMs with Ollama & llama.cpp

Source: https://www.skillbyai.com/en/local-llms/ll-modelfile

> Write a Modelfile that bakes a system prompt and parameters into a named custom model.

## A recipe for a derived model

A **Modelfile** is Ollama's plain-text recipe for building a model, similar in spirit to a Dockerfile. Each line is an instruction. `FROM` names the base: an existing Ollama model (`FROM llama3.2`) or a path to a GGUF file. `SYSTEM` sets the default system message that frames every conversation. `PARAMETER` sets defaults such as `temperature`, `num_ctx` (context window), `top_p`, `stop` sequences and `seed`. `TEMPLATE` overrides the prompt template, `MESSAGE` adds example turns for few-shot behaviour, `ADAPTER` applies a LoRA fine-tune, and `LICENSE` records terms. Running `ollama create <name> -f Modelfile` builds a new model entry. Importantly, it **reuses the base weights**; only the small configuration layers are new, so a custom model costs almost no extra disk. This is the cleanest way to give a team one shared, versioned assistant configuration.

## A DSA tutor model

Save as `Modelfile`, then build and run it.

```bash
cat > Modelfile <<'EOF'
FROM llama3.2
PARAMETER temperature 0.3
PARAMETER num_ctx 8192
SYSTEM """You are a patient DSA tutor for college students.
Explain with small examples and always state the time complexity."""
EOF

ollama create dsa-tutor -f Modelfile
ollama run dsa-tutor "How does binary search work?"
```

## Building a custom model

Creating a model layers configuration on top of existing weights. The stages, in order: base model weights -> Modelfile instructions -> ollama create -> named custom model.

![4 connected stages, left to right: base model weights, Modelfile instructions, ollama create, named custom model.](assets/figures/local-llms/section-4-map.svg) — Figure 4.1: base model weights -> Modelfile instructions -> ollama create -> named custom model

## Inspect before you override

Run `ollama show llama3.2 --modelfile` first. It prints the base model's own template, stop tokens and parameters. Copy what you need instead of guessing, because a wrong TEMPLATE silently breaks chat formatting.

**Quiz:** What does `ollama create dsa-tutor -f Modelfile` do with the base model's weights?

- [ ] It copies and re-quantizes all the weights
- [x] It reuses the existing weight layers and adds only the new configuration
- [ ] It fine-tunes the weights on the SYSTEM prompt
- [ ] It uploads them to the Ollama library

*Answer:* It reuses the existing weight layers and adds only the new configuration. Ollama stores models as layers. A derived model points to the same weight blob and adds small layers for system prompt, parameters and template. No training happens.