Ollama Just Got 2x Faster on Mac (Here's How)
Your Ollama is probably running at half the speed it could be on your Mac.
About three weeks ago, Ollama started routing certain models through Apple's MLX framework instead of the default engine. Same model, same machine, roughly 2x faster. But it's not automatic. You need to pull the right variant.
━━━━━━━━━━━━━━━━━━━━━━━━
LINKS
▸ Ollama MLX blog post: https://ollama.com/blog/mlx
▸ Companion GitHub repo: https://github.com/ravsau/ai-tutorials/tree/main/ollama-mlx-speed
▸ 1:1 help with your local AI setup: https://cloudyeti.io/meet
▸ Local LLM Playlist: https://www.youtube.com/playlist?list=PLQP5dDPLts67psUzW096fJdMNHJB16Sob
━━━━━━━━━━━━━━━━━━━━━━━━
MLX path: ~78 tokens/sec
Default path: ~36 tokens/sec
Same model. Same prompt. Same M3 Mac with 128GB.
━━━━━━━━━━━━━━━━━━━━━━━━
WHAT YOU NEED
• Ollama 0.19 or newer
• Apple Silicon Mac (M1/M2/M3/M4)
• Look for tags with nvfp4, mxfp8, or mlx-bf16
HOW TO FIND MLX MODELS
The Ollama models search page doesn't index MLX variants well yet (I've pinged their team about this). For now, check the Ollama MLX blog post or search "Ollama MLX [model name]" on the web.
━━━━━━━━━━━━━━━━━━━━━━━━
TIMESTAMPS
0:00 Ollama just got faster on Mac
0:39 Why MLX matters (vs llama.cpp)
1:02 How to find MLX-supported models
2:04 Pulling and running an MLX model
2:54 Verbose mode: 79 tokens/sec on M3
3:24 Same model without MLX (the slow path)
4:14 The speed gap, side by side
4:37 Benchmark repo to try yourself
4:50 Wrap up
━━━━━━━━━━━━━━━━━━━━━━━━
#ollama #mlx #localai #mac #applesilicon #llm #ai #opensource