TL;DR: Ollama's latest release brings MLX's custom Apple Metal kernels to CUDA, broadening GPU acceleration beyond Apple Silicon.
Summary: Ollama v0.40.5 includes a change porting MLX's custom Metal kernels to CUDA (PR #18657 by @dhiltgen), alongside an MLX version bump and OpenAI-compatible fixes for accepting custom tool calls in Responses history. The kernel port extends Ollama's MLX-based inference path to NVIDIA GPUs, which were previously limited to the Metal backend on Apple hardware.
Why it matters: NVIDIA GPU users may see faster MLX-model inference without waiting on separate backend work. Worth benchmarking MLX-format models on CUDA hardware after upgrading to v0.40.5.
Source: github_releases