Set an Ollama Output Token Limit With num_predict

Use options.num_predict to set a generation ceiling for one request. This is useful when a local tool needs a brief reply even when the model is inclined to keep writing.

Worked example

A local request with a small output budget

curl -sS http://localhost:11434/api/generate \
  -H 'Content-Type: application/json' \
  -d '{"model":"llama3.2:latest","prompt":"Describe a blue bicycle in several sentences.","stream":false,"options":{"num_predict":12}}'

Steps

  1. Start Ollama locally and choose an installed model with ollama list. Replace the model in the example if needed.
  2. Send the example request with stream set to false so you get one JSON response. Record eval_count and done_reason when the server returns them.
  3. Change num_predict from 12 to 48 and run the same request again. Compare the token counts and check whether the first reply stopped early or hit the ceiling.

What to check

  • An early stopping condition can produce fewer tokens than the limit. Check eval_count rather than counting characters in response.
  • The context length has a separate setting. Increasing num_predict does not make a long input fit into the configured context window.

Build a Modelfile for persistent defaults

References