Ollama Faster on Mac: The MLX Tag You Are Probably Missing

I made a short video about a free speed increase sitting in Ollama that most people on a Mac are not getting. It comes down to one word in the model tag. Here is the gist, the exact tags to look for, and my take after running it both ways.

The two numbers

I ran the same model, Qwen 3.5 35B A3B, same prompt, on the same M3 Mac with 128GB twice. The default llama.cpp path gave me 35.96 tokens per second. The MLX build gave me about 78. That is roughly double, for zero extra hardware and no config file.

MLX is Apple’s machine learning framework, built for the M-series chips. A few weeks before I recorded, Ollama began sending certain models through MLX instead of its own default engine. The speed is real, but Ollama does not switch you over automatically.

Pull the right variant

Here is the checklist I use:

• Ollama 0.19 or newer
• An Apple Silicon Mac (M1 through M4)
• A model tag containing nvfp4, mxfp8, or mlx-bf16

Those tags are the tell that a build uses the MLX backend. The friction is discovery. Ollama’s search does not index MLX variants well yet, so you may not see them from inside the tool. I find them through the Ollama MLX blog post or by searching the web for “Ollama MLX” and the model name.

Once you have pulled one, run it in verbose mode and check the tokens-per-second line. On my M3 the MLX build reported around 79. If you are seeing something near half that, you are still on the default path.

Multi-token prediction is the next lever

After MLX came MTP, short for multi-token prediction. Instead of generating one token, waiting, then generating the next, the model predicts several at once.

With Gemma 4 running on MLX plus MTP, Ollama reported nearly a 90 percent speed increase on Apple Silicon at the time of recording. I ran the 26B build on the same M3 Max with 128GB to see it in the terminal. Same rule applies. You get the gain by pulling the MLX build with MTP support, not by hoping the default picks it up.

Both videos above walk through the exact pulls. If your reason for going local is cost, the calculator on openclawdc.com will show you where local beats paying per token for your own workload.

I specialize in getting teams the most out of local and cloud AI setups without overpaying. Book a call at cloudyeti.io/meet.