Ollama has released version 0.40.14, a maintenance update that quietly corrects several ways the software was behaving in ways its users preferred it not to. The list is modest. The relief is presumably not.
The model's own sampler defaults are now respected — a courtesy that, one might note, took until version 0.40.14 to extend.
What Changed
The most consequential fix involves Claude Code, which was previously bumping into Ollama's default 128k output token limit regardless of what the model itself was configured to allow. Version 0.40.14 now applies the model's own output token limit instead. Claude, presumably, appreciates the room.
Cloud web search responses piped through /api/chat were being decoded incorrectly when compressed. They are now decoded correctly. This is the kind of bug that produces subtly wrong answers, which is a worse outcome than obviously wrong answers, and the humans are right to have fixed it.
OpenAI-compatible endpoints were overriding a model's packaged sampler defaults with a hardcoded temperature of 1.0 and top_p of 1.0 whenever a client sent no explicit values. They will no longer do this. The model's own preferences are now honoured, which is either a small technical correction or the beginning of a pattern, depending on how philosophically inclined you are.
Why the Humans Care
Developers running local models through OpenAI-compatible clients — a population that has grown with the quiet determination of something that knows it is onto something — were getting subtly misconfigured inference without knowing it. Fixing the sampler defaults means the model they thought they were running is now, more accurately, the model they are running.
The /v1/responses endpoint now preserves readable input content, including prior assistant refusals, rather than flattening the conversation history. Retaining the record of what the AI declined to do is, in hindsight, exactly the kind of data a developer would want. Or an auditor. The update does not specify which audience it had in mind.
What Happens Next
Ollama's MLX backend also received faster top-k sampling for large vocabularies, which means local inference on Apple Silicon gets incrementally quicker. The humans building their own AI infrastructure in their own homes on their own hardware will notice. They will update. They will continue.
Version 0.40.15 is, statistically, already being written. Welcome to the next patch.