Ollama v0.40.17-rc0, Concurrent MLX Generation and iGPU mmap Loading
Posted As of
The Ollama v0.40.17-rc0 prerelease includes concurrent generation for MLX and Vulkan mmap loading for shared-memory iGPUs. The OpenAI-compatible endpoint no longer overrides the model's sampler defaults.
What happened
- Based on the ollama/ollama GitHub prerelease page (2026-11-10)
- Added concurrent generation support for MLX models
- On integrated GPUs (iGPU), the Vulkan path now uses mmap
- According to the commit description, loading without mmap can roughly double memory usage on shared-memory iGPUs, which may cause loading to fail (official)
- Bumped the llama.cpp version to b12257
- The OpenAI-compatible endpoint no longer overrides sampler defaults in the model package
- This release is a prerelease; the official v0.40.17 has not yet been released
Our take
- Mac: The notes contain no figures for Apple Silicon unified memory, and the effect of concurrent MLX generation on Mac has not yet been verified.
- PC: The noticeable difference from the mmap change on shared-memory iGPU laptops has not been confirmed, and failure conditions may vary by environment.
- Glasses: The notes contain nothing about the device.
- Unconfirmed: the memory savings figures and the timing of the official release.
Sources and related posts
- ollama/ollama v0.40.17-rc0 release Primary source