youssofal/MTPLX

MTPLX

✅ Recently maintained

The fastest way to run Qwen 3.8 Flash Next, Qwen 3.8 27B and Ternary Bonsai 2 27B on a Mac: 125 tok/s in OpenCode on an M5 Max, and a 27B model on 16 GB Macs. Native MTP speculative decoding on Apple Silicon, exact at any temperature. OpenAI and Anthropic compatible local server.

2,669Stars
210Forks
+327-day stars

Python · Apache-2.0 · Last push 2026-10-31

Official repository & installation ↗
apple-siliconlocal-aimlxmtpopenai-compatibleqwenspeculative-decodinganthropic-compatibleinference-enginemetal

GitValue · Value evidence

Unscored · Evidence coverage 0%

Provisional rubric; documentation and runtime testing are separate. Missing evidence does not imply low value.

Star history

2026-04-272026-11-03
View historical data
DateStars
2026-10-052,513
2026-10-062,518
2026-10-072,529
2026-10-082,537
2026-10-092,546
2026-10-102,554
2026-10-112,559
2026-10-122,568
2026-10-132,573
2026-10-142,578
2026-10-152,582
2026-10-162,587
2026-10-172,590
2026-10-182,594
2026-10-192,603
2026-10-202,610
2026-10-212,617
2026-10-222,621
2026-10-232,622
2026-10-242,625
2026-10-252,628
2026-10-262,634
2026-10-272,643
2026-10-282,648
2026-10-292,652
2026-10-302,653
2026-10-312,658
2026-11-012,661
2026-11-022,667
2026-11-032,669

Get the project

git clone https://github.com/youssofal/MTPLX.git

Cloning downloads the source. For dependencies and installation, follow the project README.

Related projects & alternative candidates

Related domains do not imply feature parity. Review official documentation and licenses.

omlx

LLM inference server for Apple Silicon with continuous batching and SSD caching, managed from the macOS menu bar.

★ 23,304