Ollama 0.40.7 fixes Responses conversations for Codex and lets more draft models speed up replies

Ollama 0.40.7 is a fix release for people who drive Ollama from coding tools: it repairs two faults in its OpenAI-compatible Responses interface, one of which showed up with Codex , lets more draft models speed up replies through speculative decoding , and lists a new decision model, Perplexity's pplx-decider. The release page carries the date 18 October, 00:28 , which is when the page was first created; GitHub keeps that date when a release is made final later. The installers it offers were uploaded on 21 October at 03:53 UTC , and its changelog line runs from v0.40.6 to v0.40.7.

What changed

Responses conversations that no longer break.The Responses interface is the OpenAI-compatible way coding tools such as Codex talk to Ollama; each request carries the conversation so far. Continuing such a conversation no longer fails when you switch between models with and without thinking — the step in which a model reasons before it answers — or when the history contains another provider's compaction items. Compaction is how a tool shortens a long history to save room, and each provider leaves its own kind of items behind.

Tool calls in the right order.In streaming Responses requests, where the answer arrives piece by piece, client tool calls — the page's example is Codex with web search — could be emitted before open reasoning items, so tool call events arrived out of order or twice. That is fixed.

More draft models for speculative decoding.Speculative decoding is a way to get answers faster: a small, quick "draft model" guesses the next few words, and the large model checks the guesses in one pass and keeps the ones it agrees with. More draft models now work this way, because llama-server works out the draft type itself instead of requiring Ollama to identify it.

A new decision model.pplx-decider is a 27B decision model from Perplexity, fine-tuned from Qwen3.8-27B. If decision models are new to you, our itemOllama 0.35 adds decision models that answer with choices, probabilities and scores instead of text explains what they return and how to question them.

The rest of the list is smaller:

• ollama show now reports which renderer and parser a model uses and where its chat template comes from, and the API exposes the same values in new renderer, parser and template_sourcefields. The chat template is the pattern that turns your conversation into the exact text the model reads; broadly, the renderer builds that text and the parser reads the model's output back into an answer, its thinking and its tool calls.
• Requests routed to remote models — the cloud tags that run on ollama.com rather than on your machine — are now cancelled there when the client disconnects, instead of continuing to run. Our itemOllama 0.40.5 brings NVIDIA's Nemotron 3.5 Super and its smaller sibling, Bolt covers those cloud tags.
• ollama launch claude now clamps CLAUDE_CODE_AUTO_COMPACT_WINDOWto Claude Code's supported range, avoiding a "window too large" warning.
• llama.cpp and MLX are updated.

What it means for you

If you use Codex, or any tool that talks to Ollama through the Responses interface, update.If a conversation stopped with an error after you switched between a thinking and a non-thinking model, or after you carried a history over from another provider, that is the first fix. If you saw the same tool call twice, or out of order, while Codex searched the web, that is the second.

If you use speculative decoding,a draft model that did not work with Ollama before may work now. Try the pair again after updating, and time the same prompt with and without the draft model on your own machine: how much it saves depends on how often the draft guesses right.

If you start Claude Code with ollama launch claude,the "window too large" warning goes away. To clamp is to bring a value outside the range back to its nearest edge: if you setCLAUDE_CODE_AUTO_COMPACT_WINDOW yourself outside Claude Code's range, the value in use is now that edge, not the one you typed.

If you want to try pplx-decider,27B is its size ; the page gives no pull command and no memory figure. Our guideWhich Machine Runs a Model Locally explains how a model's size decides whether it fits on your machine.

How to update.The release's list of files carries an installer for each system, all uploaded on 21 October. On Windows, download and runOllamaSetup.exe; on a Mac,Ollama.dmg; on Linux, theinstall.shscript from the same list. Then open a terminal and typeollama -v: it prints the version you have installed, which should now be 0.40.7. If it shows an older number, quit Ollama completely, start it again and run it once more.

The source

The release page on GitHub, Ollama v0.40.7, and its list of downloadable files with their upload times.