MTPLX vs oMLX

🏎️ MTPLX vs oMLX: the difference in local LLM inference speed on Mac — 2.8–6.4×

The comparison of two MLX servers — MTPLX and oMLX — came out on 03 October 2026. The runs were done on a MacBook Pro with an M4 Max chip, the test model was Qwen3.8-27B. The claimed difference in local LLM inference speed on Mac: 3–6×. According to the run results, MTPLX finished the task 2.8–6.4 times faster than oMLX.

The highlight of MTPLX is multi-token prediction (MTP).

The reason for the comparison was a conversation with ChatGPT: it recommended oMLX as the faster and higher-quality option on M4 Max. In practice, the advice did not hold up 🧪

This is a continuation of the series. The first part of the article was titled "A local coding agent on Mac: MTPLX + pi + Qwen3.8-27B" — there, the coding agent pi was hooked up to a local Qwen3.8-27B via MTPLX.

Test nuances: on hidden runs, the MTPLX and oMLX setups came out even. The model builds are different, and each mode on oMLX was run only once. The full breakdown is in the article on habr.com