The Pulse

Ollama v0.40.7 RC Fixes Stale Reasoning Settings in Responses

Ollama’s v0.40.7-rc0 release changes how its Responses endpoint handles saved reasoning-effort settings when a request targets a model without thinking support. The release also lists updates to llama-server’s draft speculative type inferen

AI.info Team · Written with AI · 2026-10-18

“Switching a Responses client from a thinking model to llama3.2 can carry reasoning.effort=high into the next turn.”

hoyyeva, author of the pull request

That setting could make Ollama reject a request after a user changed models. Ollama’s v0.40.7-rc0 release, published October 18, changes the OpenAI-compatible Responses endpoint so a saved reasoning-effort value does not cause that failure when the selected model lacks thinking support.

Why a saved effort setting caused an error

The problem arises when a client moves from a model with thinking levels to one without them. The client can retain reasoning.effort=high and send it with a later request for a model such as llama3.2. In the pull request, author hoyyeva said the installed ChatGPT backend sends the saved effort even when the selected model offers no thinking levels, and Ollama rejects the turn with a “does not support thinking” error.

The pull request’s discussion gives examples of how the failure appears: a non-streaming request returns an HTTP 400 error, while a streaming request emits response.failed. In the reported reproduction, omitting the effort setting or explicitly setting think: false succeeds. The reported issue concerns the carried-over setting in Responses requests, rather than a new reasoning capability being added to models that lack thinking support.

The change applies an existing fallback for models without thinking capability to effort-derived Responses settings. In effect, Ollama ignores the carried-over effort when the model cannot use it, rather than treating the leftover setting as a reason to fail the request. The fix does not give a non-thinking model a new reasoning mode.

Web-search follow-ups get the same resolved setting

The patch also addresses a second failure path: a Responses request that uses web search may succeed at first, then fail when Ollama sends a follow-up request to the model. In a review comment on the pull request, Parth Sareen described how the initial inference and search could succeed, but the follow-up could retain the original high-effort setting and trigger the unsupported-thinking error.

The pull request describes the underlying issue as a setting that applied to the initial request but was absent from the follow-up request’s context. Its change reuses the resolved thinking value for web-search follow-ups, so the later model request receives the same resolution. The discussion says that this path was reproduced with a non-thinking model, reasoning.effort: "high" and web_search; the follow-up failed, while omitting the effort or setting think: false worked.

The pull request says explicit think overrides keep their precedence and validation, while boolean and named-level resolution remains unchanged. It also says native API and Chat Completions behavior are unchanged. The adjustment is therefore scoped to Responses effort handling and the web-search follow-up flow described in the change.

Three changes in the v0.40.7-rc0 release

The reasoning-effort adjustment is one of three changes listed in the release candidate. Ollama also says llama-server can infer the draft speculative type, and its MLX inference support adds PPLX Decider v1.2 support.

The release is marked as a pre-release on GitHub, so v0.40.7-rc0 is a release candidate, not a final version. For developers using the Responses API, its practical effect is narrower than a new reasoning control: saved effort settings should no longer break turns on models that do not support thinking, including the follow-up request after web search.