Ollama 0.40.7 adds Perplexity's pplx-decider model, fixes OpenAI-compatible streaming bugs

Ollama 0.40.7 adds pplx-decider, a 27B decision model from Perplexity based on Qwen3.8-27B. The release fixes conversation failures when switching between thinking and non-thinking models in OpenAI-compatible Responses, resolves misordered tool call events in streaming, and cancels remote model requests on client disconnect. The ollama show command and API now expose renderer, parser, and template_source metadata. Speculative decoding gains broader draft model support through automatic type inference in llama-server. Underlying llama.cpp and MLX libraries are updated.

What's new in Ollama 0.40.7

• New model: pplx-decider (27B, Perplexity, fine-tuned from Qwen3.8-27B)
• OpenAI Responses fixes: conversation switching between thinking/non-thinking models; handling of other providers' compaction items in history
• Streaming fix: prevents client tool calls (e.g. Codex web search) from emitting before open reasoning items, which caused misordered/duplicate events
• Remote request cancellation: disconnecting clients now cancel running requests on ollama.com instead of leaving them consuming resources
• New metadata exposure: ollama show and API return renderer, parser, and template_source fields for model introspection
• Speculative decoding: llama-server auto-infers draft model type, removing Ollama-side identification requirement
• Claude Code integration: ollama launch claude clamps CLAUDE_CODE_AUTO_COMPACT_WINDOW to supported range
• Engine updates: llama.cpp and MLX refreshed

Why it matters

The automatic draft model inference for speculative decoding lowers the barrier for users experimenting with faster local inference. Exposing renderer and parser metadata improves transparency for developers building on Ollama's API.