Ollama's OpenAI-compatible endpoints now keep each model's sampler defaults (November 3, 2026)

What changed

Ollama v0.40.14 (released November 3) changes how its OpenAI-compatible endpoints set sampling. The release notes say these endpoints no longer override a model's packaged sampler defaults. A client that sends no temperature or top_p now gets the model's own values instead of 1.0, the same as /api/chat.

The same release also includes these changes:

/v1/responses keeps readable input content, including earlier assistant refusals.
Cloud web search responses on /api/chat now decode correctly.
ollama launch claude applies the model's output token limit.

Prerequisites

Ollama v0.40.14 or later, running locally
No key is needed for local models

Steps

Upgrade Ollama to v0.40.14 or later.
Find every call your code makes through Ollama's OpenAI-compatible endpoints that leaves out temperature or top_p.
If your output depended on the old 1.0 values, send temperature and top_p explicitly in each request.
If you want the model's own defaults, leave both fields out.

Test it

Send the same prompt twice: once with no sampler fields, and once with temperature set explicitly. Compare how much the outputs vary, then compare the first result with the same request sent to /api/chat.

Not confirmed

The release notes do not list each model's packaged default values. They also do not say whether the change covers every OpenAI-compatible route or only some.