Ollama makes MLX the default runtime on Apple Silicon
By Leon Mallett
Originally announced by Ollama (GitHub releases) . We link the primary source so you can read it for yourself.
The news
Ollama, one of the most common ways to run models locally, has made MLX the default runtime on Apple Silicon in its v0.40.0 release. On a Mac, model architectures that the MLX runtime supports now run on MLX automatically, rather than through Ollama’s usual llama.cpp path, with no manual configuration. The release also broadened MLX support to more recent models, including the Qwen3.x line and Gemma 4.
A note on dating: the version and change are confirmed on Ollama’s GitHub releases, which places v0.40.0 in late September 2026, though some secondary coverage gave an early-October date. Confirm the exact release date on the releases page before relying on it.
Why it matters
MLX is Apple’s machine-learning framework, built for Apple Silicon and its unified memory architecture. Running a model through MLX rather than a more general backend can make better use of that memory and the Apple GPU, which on a Mac can mean faster inference and more headroom for context at the same memory budget. Because a large share of local AI happens on Macs, a default that picks the Apple-native path without the user having to know about it removes a real piece of friction. It is the kind of change that quietly shifts what is practical on hardware people already own.
The history
This continues a line of Apple Silicon work in Ollama through 2026; we covered an earlier Apple Silicon improvement in August. The direction is consistent: treat Apple Silicon as a first-class local-AI platform rather than a port of the CUDA path.
The practical detail
The change applies to model architectures the MLX runtime supports, so the benefit depends on whether the specific model you run is covered; models outside that set still use the existing path. As with any runtime change, it is worth checking your own model and workload rather than assuming a uniform speed-up. If you are weighing Ollama against the alternatives on a Mac, our guide on LM Studio versus Ollama and the wider running models topic cover the trade-offs.