Intel / repo

ollama v0.40.7

Source: github.com
Author: github-actions[bot]
Date: Published Oct 18, 2026

Context

The headline addition in Ollama v0.40.7 is pplx-decider, a 27B decision model from Perplexity fine-tuned from Qwen3.8-27B (release notes). Perplexity's Decider models are built to answer typed questions, such as yes-or-no or pick-a-category, with probabilities instead of free-form prose, as the coverage of their arrival on Vercel's AI Gateway explains. The Ollama notes do not say which Decider version this is or how its question types are exposed through Ollama.

The most consequential fix concerns remote models. Ollama can route a request to a model running on ollama.com instead of on the local machine, and until this release such a request kept running there after the client disconnected. It is now cancelled when the client goes away.

Several fixes target clients that use Ollama's OpenAI-compatible Responses API, such as OpenAI's Codex. Continuing a conversation no longer fails when it switches between models with and without thinking, or when its history contains another provider's items. requests no longer emit client tool calls, for example Codex's web search, ahead of open reasoning items, which had produced misordered or duplicate tool call events.

Key takeaways · AI-distilled

Ollama v0.40.7 now cancels requests routed to remote models on ollama.com when the client disconnects; before this release those requests kept running after the caller was gone.
More draft models now work with in Ollama, because llama-server infers the draft model type automatically instead of requiring Ollama to identify it.
`ollama launch claude` now clamps CLAUDE_CODE_AUTO_COMPACT_WINDOW to the range Claude Code supports, which avoids a "window too large" warning.
The Responses API fix also covers conversation histories that contain another provider's compaction items, so continuing those conversations in Ollama no longer fails.

Terms in this piece · Glossary

context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.

Read the source github.com