Ollama MLX Nearly Doubles LLM Speed on Your Mac: 58 to 112 tok/s, and Your Old Models Get None of It | by Kashif Mehmood | Jul, 2026
16 hours ago
Ollama’s decode speed on Apple Silicon went from 58 tokens per second to 112 in one release. Same machine, same size of model, same ollama run. Version 0.19, tagged on 27 March 2026 and announced three days later, swapped the engine underneath from llama.cpp’s Metal backend to MLX, Apple’s own machine learning framework.
If you updated that week, ran the models you already had pulled, and timed them, you measured exactly zero of it.
That is not a bug. Ollama now runs two inference engines side by side, and the file format of each model decides which one you get. Every GGUF file on your disk still goes to llama.cpp. Only safetensors-format models, like the NVFP4-quantised Qwen3.5 build the release shipped with, touch MLX at all. The fastest local-inference upgrade Macs have seen this year is gated behind a pull most people never ran.
The switch was thirteen months in the making. Ollama maintainer Daniel Hiltgen opened a draft MLX Go backend on 14 February 2025, back when it could only load llama3 fp16 and the KV cache drifted after a few forward passes. It shipped as a preview in March 2026 and hit the Hacker News front page at 648 points. This piece is about what actually changed, why…
