The Pulse
Ollama 0.40 Makes MLX the Default for Supported Mac Models
Ollama’s v0.40.0-rc2 release candidate automatically routes supported model architectures to MLX on Apple Silicon, while the project says it is still testing and enabling additional models. The change builds on Ollama’s earlier MLX preview
AI.info Team · Written with AI ·
Ollama’s v0.40.0-rc2 makes MLX the automatic route on Apple Silicon for model architectures its MLX runtime supports, but the release does not promise coverage for every model. That distinction puts a limit on the new default: Mac users get a simpler path where support exists, while Ollama says it is still testing and enabling additional models.
MLX becomes the automatic choice for compatible models
Released on September 25, the v0.40.0-rc2 notes say model architectures supported by the MLX runtime will automatically run on MLX on Apple Silicon devices. The example uses ordinary commands—ollama pull qwen3.8 followed by ollama run qwen3.8—without specifying an MLX suffix. Ollama’s wording makes the condition explicit: the automatic route applies to architectures the runtime supports. The notes add that the project will test and enable additional models during the prerelease.
That caveat matters more than the shorthand “MLX by default.” The notes describe a change in how compatible models are selected, not a claim that every model in Ollama’s library already runs through MLX.
A shift in how Mac users select the runtime
Ollama introduced its MLX-powered Apple Silicon preview in March, describing MLX as Apple’s machine-learning framework and citing the chips’ unified memory architecture. By June, Ollama’s instructions for running Gemma 4 through that engine used the tagged model name gemma4:12b-mlx. The v0.40 release candidate’s plain-name Qwen example signals a change in the default selection for supported architectures, rather than the first arrival of MLX in Ollama.
Ollama’s earlier material sets that change in context. In its March 30 preview, the company said its Apple Silicon engine was built on MLX to use unified memory and advised users to have a Mac with more than 32 GB of unified memory. It also said it was working to expand supported architectures and planned an easier import path for custom models fine-tuned on supported architectures. Those statements describe the rollout as incremental, consistent with RC2’s note that additional models would be enabled during prerelease.
The distinction also leaves room for Ollama’s other model path. In a June release post, the company described improved GGUF compatibility through llama.cpp as an addition to its MLX engine on Apple Silicon. The v0.40 notes do not say that the GGUF path disappears or that every model changes runtimes; they announce automatic MLX use only for architectures the MLX runner supports.
RC2 is still a test release
Ollama labels v0.40.0-rc2 a pre-release. On the project’s releases page, v0.35.1 carries the “Latest” label, while v0.40.0 is marked “Pre-release,” so the MLX default described in RC2 is not yet presented there as the stable release.
For developers evaluating it, the release notes offer a quick test: pull and run a supported model under its regular name, then confirm the runtime behavior for that model and build. The unresolved boundary is the one Ollama identifies itself—how many more architectures it will enable before v0.40 leaves prerelease. GitHub still marks v0.35.1 as the latest stable release.