What is a Modelfile in Ollama?

A Modelfile is a plain-text recipe for a custom model. It sets a base model with FROM , an optional system prompt with SYSTEM , and default options with PARAMETER . You build it into a model with ollama create .

How do I set a permanent system prompt for a model?

Add a SYSTEM instruction to the Modelfile, then run ollama create . Every run of that model then starts with the same role, so you do not have to repeat the instructions in each prompt.

How do I make a model's output deterministic?

Set PARAMETER temperature 0 and a fixed PARAMETER seed . Temperature zero removes randomness, and a fixed seed makes repeated runs return the same result for the same prompt.

Do I need to rebuild after editing a Modelfile?

Yes. Editing the Modelfile does not change an existing model. Run ollama create -f Modelfile again so the new settings are baked into the model.

Build / AI / Ollama

Modelfile Blueprints

Write a Modelfile to set a base model, bake in a system prompt, and lock default parameters.

TL;DR

FROM sets the base model every Modelfile builds on.
SYSTEM bakes a persistent persona into the model.
Set defaults like PARAMETER temperature 0 for repeatable output.

Jump to section

Modelfile Basics
Common Parameters
Build And Run
Beyond The Basics
Tips
Warnings
In Practice
FAQ

Modelfile Basics

FROM

Set the base model or weights to build on.

FROM llama3.2

SYSTEM

Bake a persistent system prompt into the model.

SYSTEM """You are a terse SQL tutor."""

PARAMETER

Set a default runtime option for every run.

PARAMETER temperature 0.2

Comments

Lines starting with a hash are ignored.

# This is a comment in a Modelfile

Common Parameters

temperature

Control randomness; 0 is deterministic, higher is creative.

PARAMETER temperature 0

num_ctx

Set the context window size for the model.

PARAMETER num_ctx 8192

stop

Add a stop sequence that ends generation.

PARAMETER stop "<|im_end|>"

seed

Fix the random seed for reproducible output.

PARAMETER seed 42

Build And Run

ollama create

Build a named model from the Modelfile.

ollama create sql-tutor -f Modelfile

Run It

Start the custom model like any other.

ollama run sql-tutor

Rebuild After Edits

Re-run create so edits take effect.

ollama create sql-tutor -f Modelfile

Inspect It

Show the resolved Modelfile of a model.

ollama show sql-tutor --modelfile

Beyond The Basics

ADAPTER

Apply a LoRA adapter on top of the base.

ADAPTER ./lora-adapter.gguf

LICENSE

Embed a license string in the model.

LICENSE """MIT License"""

TEMPLATE

Override the prompt template, covered next chapter.

TEMPLATE """{{ .Prompt }}"""

Tips

Put a clear SYSTEM prompt in the Modelfile so every run starts with the same role, without re-typing instructions each time.
Set PARAMETER temperature 0 when you need deterministic output, such as extraction or classification tasks that must be repeatable.

Warnings

ollama create rebuilds the model each time; re-run it after editing the Modelfile, or your changes will not take effect.
A very low temperature can make a model repetitive or terse; test values between 0 and 0.7 for your task.

In Practice

Write a Modelfile with a system prompt and zero temperature, build it, then run repeatable output.

The heredoc writes a Modelfile with a system role and fixed seed.
temperature 0 plus a fixed seed make output repeatable.
ollama create bakes all of it into a named model.
Running the model needs no extra flags; the behavior is built in.

# Write the Modelfile
cat > Modelfile <<'EOF'
FROM llama3.2
SYSTEM """Extract dates as YYYY-MM-DD only."""
PARAMETER temperature 0
PARAMETER seed 42
EOF

# Build the custom model
ollama create date-bot -f Modelfile

# Run it; behavior is baked in
ollama run date-bot "Meeting next Tuesday"