For a long time, Ollama was my default recommendation for running local language models. It remains one of the easiest options available:
ollama run qwen3I still recommend it for Windows and Linux—and for anyone who wants the simplest possible setup. On my Mac, however, I now prefer oMLX.
Built around MLX
MLX is Apple’s machine-learning framework for Apple Silicon. It is designed to take advantage of the unified memory in M-series Macs, making it a natural foundation for local AI.
Ollama now has an MLX engine, with improved performance, caching, and model-specific optimizations. That is a significant improvement, but its MLX support still depends on Ollama adding support for each model architecture.
oMLX takes the opposite approach: it is built around MLX from the start. It works with compatible MLX models from Hugging Face, giving me faster access to newly released models without waiting for them to appear in Ollama’s library.
That does not mean it supports every Hugging Face repository—the model still needs a compatible MLX format and architecture—but the selection is broad.
More than a model runner
oMLX feels like a complete Mac application rather than just a command-line server. It includes:
- A native macOS menu-bar app
- A dashboard for downloading and managing models
- LLM, vision, embedding, and reranker support
- Multi-model serving and automatic memory management
- OpenAI- and Anthropic-compatible APIs
- Tool calling and optional MCP integration
- In-memory and persistent SSD caching
The tiered cache is particularly useful for coding agents. Frequently reused prompt data can remain in memory or be restored from SSD instead of being processed again from scratch.
Where Ollama still wins
Ollama remains the better choice for simplicity and cross-platform support. It runs on macOS, Windows, and Linux, has a large ecosystem, and makes downloading and running a supported model almost effortless.
oMLX is much more specific: it requires Apple Silicon and macOS 15 or newer. It also exposes more settings, which is great for control but can feel more complicated.
My current recommendation is therefore simple:
- Windows or Linux: use Ollama.
- Mac and maximum simplicity: use Ollama.
- Apple Silicon Mac and the best MLX-native experience: try oMLX.
Ollama’s MLX engine is improving quickly, but oMLX still feels more comprehensively built for the Mac. For now, it is my favorite way to run local models.
Leave a Comment