Ollama

KnowledgeGraphs

GraphAI

Context window and num_ctx

The context window (aka context length) is the maximum number of tokens a model can consider at once, it’s determined by the model architecture. For example, if a model has a context window of 8192 tokens, it can process up to 8192 tokens of input and output combined. The num_ctx on the other hand is the Ollama-specific parameter that sets the active context length for a given run. This means that if you set num_ctx to 8192, the model will utilize the full context window available for that model during inference. If you set it lower, say 4096, the model will only use up to 4096 tokens of context, even if the model supports more.

You can view the context window using Ollama’s CLI command:

ollama show qwen2.5:14b

This renders:

 Model
    architecture        qwen2
    parameters          14.8B
    context length      32768
    embedding length    5120
    quantization        Q4_K_M

  Capabilities
    completion
    tools

  System
    You are Qwen, created by Alibaba Cloud. You are a helpful assistant.

  License
    Apache License
    Version 2.0, January 2004

This tells you that Qwen2.5:14b has a context length of 32768 tokens.

Custom Model

You can alter the context window of any Ollama model by creating a custom model with a different num_ctx parameter.

Create a Modelfile with the following content:

# My custom Modelfile to increase context length
FROM qwen3:8b

# Set the context window size (e.g., 4096, 8192, 16384, etc.)
PARAMETER num_ctx 8192

and save this as (for example) Modelfile.custom.

Then build the custom model using the Ollama CLI:

ollama create qwen3:8b-8k -f Modelfile.custom

This will fetch the based model if necessary and create a new model qwen3:8b-8k with an 8192 token context window.

You can also export a model alter the model file if you prefer:

ollama show --modelfile qwen2.5:14b > ./qwen2.5-14b-model

Edit this file to change the num_ctx parameter or set

PARAMETER num_ctx 8192

and build as before:

ollama create qwen2.5:14b-8k -f ./qwen2.5-14b-model