Ollama v0.40.7

Summary

Ollama v0.40.7 is a small release with a few useful changes. ollama show reports where a model's chat template comes from, more draft models work for speculative decoding, and ollama launch claude keeps Claude Code's auto-compact window in range. It also adds pplx-decider, Perplexity's 27B decision model.

Key points

• ollama show prints each model's renderer, parser, and chat template source. The API exposes them as renderer, parser, and template_source. This helps when you debug tool calling or thinking output from a local model.
• Speculative decoding: llama-server now infers the draft type automatically, so more draft models work.
• ollama launch claude (which starts Claude Code against Ollama) clamps CLAUDE_CODE_AUTO_COMPACT_WINDOW to the range Claude Code supports.
• Requests to ollama.com cloud models are cancelled when the client disconnects. Responses-API conversations no longer fail when switching between thinking and non-thinking models.
• New model: pplx-decider, a 27B model from Perplexity fine-tuned from Qwen3.8-27B for decision tasks.
• Updated llama.cpp and MLX.

Why it matters

These are small changes for anyone who runs Claude Code or other agents against local models through Ollama. The template-source fields make it easier to see why a local model formats tool calls wrongly.