Ollama Just Made Every Mac 2x Faster at Running AI
The local AI inference landscape just shifted overnight. Ollama 0.19 ships with native MLX support, meaning every Mac with Apple Silicon just got dramatically faster at running AI models locally.
What Changed
Ollama --- the most popular tool for running AI models on your own hardware --- previously used llama.cpp as its inference engine. It worked, but it wasn't optimized for Apple's unified memory architecture.
Version 0.19 swaps in MLX, Apple's own machine learning framework built specifically for Apple Silicon. The results aren't subtle:
- 1.6x faster prefill (how quickly the model processes your prompt)
- ~2x faster decode (how fast it generates responses --- 58 tokens/sec to 112 tokens/sec)
- Lower memory usage with intelligent KV cache checkpointing
Awni Hannun, the lead on Apple's MLX project, called it the moment he'd been waiting for since MLX was first open-sourced.
Why This Matters for You
If you're running a MacBook Pro, Mac Mini, Mac Studio, or any Apple Silicon machine, you just got a free performance upgrade for local AI. No hardware changes. No new subscriptions. Just update Ollama.
This is particularly significant for three use cases:
Personal AI assistants that run entirely on your machine. Tools like OpenClaw can now orchestrate local models with response times that feel closer to cloud APIs.
Coding agents like Claude Code and OpenCode that need fast local inference for code completion and analysis. The 2x decode speedup means less waiting between suggestions.
Privacy-sensitive work where you can't send data to cloud providers. Faster local inference makes the privacy tradeoff much less painful.
The Bigger Picture
This effectively kills the debate about whether you need specialized inference servers (like vLLM or SGLang) for Apple Silicon. The argument was always that Ollama wrapped llama.cpp, which wasn't optimized for unified memory. Now that argument is gone.
For anyone building an AI workflow on a Mac, the stack just got simpler: install Ollama 0.19, pull your models, and go.
The unified memory architecture in Apple Silicon was always theoretically ideal for LLM inference --- large models can sit entirely in shared CPU/GPU memory without transfer bottlenecks. MLX was built to exploit that. Now the most popular local inference tool uses it natively.
What to Do
- Update Ollama to version 0.19+ (preview available at ollama.com)
- Your existing models work as-is --- no re-downloading needed
- Expect the biggest improvements on larger models where memory bandwidth matters most
The gap between local and cloud inference just got smaller. And unlike cloud APIs, this performance improvement costs you nothing.