Built on MTPLX
| Project | What it takes from MTPLX | Their words |
|---|---|---|
| oMLX 21,371 stars | Lightning MTP runs on MTPLX's verify-shape Metal kernels. Loads MTPLX-format packs since 11 May 2026, imports MTPLX side-car MTP heads since 8 Aug 2026. | "Lightning MTP's verify-shape Metal kernels are powered by MTPLX by Youssof Altoukhi, which also inspired the depth-k pipeline." |
| mlx-serve 1,105 stars | The verify-width split-K quantized matmul family and the M5 NAX tensor-ops tile. | "Powered by MTPLX by Youssof Altoukhi." |
| edgequake-llm | Ships an MTPLX provider. | |
| pi-mtplx | Runs the Pi coding agent against MTPLX with automatic model discovery and live token-speed monitoring. | |
| hipdash | A live dashboard for MTPLX: per-request metric history, session bank charts, benchmark runs. | |
| llm_context_benchmarks | Ivan Fioravanti's context-length benchmark suite ships mtplx_benchmark.py. |
Coverage
- MTPLX: 2x Faster Local LLMs on Apple Silicon, Wired Into DevoxxGenie, DevoxxGenie blog. "The speedup is real, the memory cost is zero, and the output distribution is provably unchanged."
- I Got 18 tok/s From a 27B Model on a MacBook: MLX + Native MTP Speculative Decoding on Apple Silicon, 19 May 2026. Independent same-machine run: llama.cpp MTP 10.5 tok/s, MTPLX depth 3 18.3 tok/s on Qwen 3.6 27B.
- GitHub monthly Trending, 1 to 4 September 2026, 826 stars in the month.
The numbers
| Metric | Value | As of |
|---|---|---|
| GitHub stars | 2,004 | 3 Sep 2026 (star 1,000 on 12 Jul, star 2,000 on 3 Sep) |
| GitHub forks | 148 | 3 Sep 2026 |
| Hugging Face downloads, all MTPLX packs, 30 days | 138,089 | 3 Sep 2026 |
| Hugging Face downloads, Qwen 3.8 packs, 30 days | 119,875 | 3 Sep 2026 |
| Hugging Face downloads, Qwen 3.8 27B Optimized Speed, 30 days | 61,275 | 3 Sep 2026 |
| Tagged releases since 2 May 2026 | 52 | v0.1.0-preview to v2.10.2 |
| First commit to first public release | 5 days | 27 Apr to 2 May 2026 |
Press kit
Name. MTPLX, all capitals, in prose. The CLI command is lowercase mtplx.
One line. MTPLX is a free, open source Mac app and CLI that runs local LLMs twice as fast on Apple Silicon with native multi-token-prediction speculative decoding, exact at any temperature.
One paragraph. MTPLX runs the multi-token-prediction heads that models such as Qwen 3.8 ship with: the model drafts several tokens ahead of itself, verifies the block in one batched forward pass, and commits tokens through exact rejection sampling with residual correction, so the output distribution is the model's own. There is no second draft model and no greedy shortcut. It serves an OpenAI-compatible and an Anthropic-compatible API on one port for OpenCode, Pi, Claude Code, Cline and Open WebUI, and keeps long agent sessions warm with a prefix cache that runs alongside speculation. Youssof Altoukhi created it in April 2026. Apache-2.0.
Creator. Youssof Altoukhi (GitHub youssofal, X @Youssofal_, Hugging Face Youssofal). Company: YOYO STUDIOS INC.
Attribution for derived work. "Powered by MTPLX by Youssof Altoukhi", linked to mtplx.com.
Assets. Wordmark (PNG), dashboard image (1600 x 1262), app icon.
Contact
Bug reports and benchmark replications: GitHub issues. Everything else: @Youssofal_ on X.