Ollama v0.40.7: PPLX Decider v1.2 support and speculative decoding improvements

The Ollama v0.40.7 release of October 18, 2026 adds support for the PPLX Decider v1.2 model for MLX, allows the effort parameter in the Responses API for models without reasoning, and lets llama-server automatically detect the draft model type for speculative decoding.

What it means

For those who run models locally via Ollama, the update brings three changes. First, on Apple Silicon machines you can now work with PPLX Decider v1.2 — a model that chooses between different response options. Second, the effort parameter in the Responses API is no longer limited to reasoning models: it can also be passed for regular ones, although the developer does not specify exactly how it affects generation in that case. Third, llama-server now detects the draft model type for speculative decoding on its own — previously this had to be specified manually. The update is narrowly targeted and mainly affects those who use the Responses API or experiment with speculative decoding. You can follow news about tools for running neural networks locally in the CPU3D news feed.