How to Configure a Local Model in Ollama: Parameters for Full Control

Ollama provides access to powerful language models that run entirely locally on your computer. But simply running a model is only half the battle. Tuning its parameters lets you turn one model into several different tools depending on the task.

Why Tune Parameters at All?

Ollama ships with reasonable default values. But if you need a model for creative writing, it will work differently than for code. Temperature affects the originality of responses, context determines how far back the model looks in conversation history, and performance parameters decide whether the model will run on CPU or GPU.

Main Parameters

• temperature — how creative the model will be (0 = predictable, 1 = creative). For code, set 0–0.3; for ideas, 0.7–1.0
• top_p — diversity in word selection (probability filter). Usually 0.8–0.95
• top_k — the number of options from which the model selects the next word. 40–50 is a good compromise
• num_ctx — context size in tokens (what the model remembers from the conversation). 2048–4096 for most tasks, up to 8192 if you have the memory
• num_thread — how many CPU cores to use. More ≠ always faster; experiment
• num_gpu — how many layers of the model to send to the GPU. Speeds things up if you have a GPU

When to Change Which Settings

For chat (assistant, questions), set temperature = 0.7, top_p = 0.9, and num_ctx = 4096. The model will be natural but won't lose conversation context. For coding — temperature = 0.2, num_ctx = 2048, num_thread = 8. For creative content (ideas, scripts) — temperature = 0.9, top_p = 0.95, num_ctx = 3072.

"The best way to find optimal parameters is to experiment and look at the results," advice from the Ollama community.

You can export your settings to a Modelfile right now and share them with others. Ollama reads parameters from this file when loading a model.

What This Means

Ollama turns local models into a tool, not a black box. You control not just the data (it stays on your machine) but also the model's behavior. This is especially important for companies concerned about privacy and for developers experimenting with LLMs without cloud services.