ollama v0.40.14

Ollama v0.40.14 changes what happens when a client calls Ollama's OpenAI-compatible endpoints without setting temperature or top_p, the two sampling settings that control how much randomness goes into choosing each next token (release notes). Until this release, those endpoints filled the gap with 1.0, overriding the sampler defaults packaged with the model. Such clients now get the model's own values, which is what Ollama's native /api/chat endpoint already did. An app that reaches the same model through both interfaces should now get the same sampling behavior, and an app that relied on the old 1.0 default may see its outputs change after upgrading.

Several other fixes in the release address the same kind of mismatch between Ollama's compatibility layers and the models behind them. The ollama launch claude command now applies the model's own output token limit for Claude Code, so cloud models no longer run into a 128k default cap. The /v1/responses endpoint now keeps readable input content, including earlier assistant refusals, instead of flattening it. The ollama launch codex command checks that ChatGPT routing works before configuring and reports a clear error when it is unavailable, and Ollama's app can now retry ChatGPT disconnect failures instead of leaving the connection broken. The remaining changes fix decoding of cloud web search responses on /api/chat, including compressed ones, and speed up top-k sampling on MLX for large vocabularies (full changelog).