youssofal/mtplx is a Python app and command-line tool that runs local models on Apple Silicon. The latest release is v2.12.2. The provided records do not show what was added or changed in this version. The tool performs speculative decoding using the multi-token prediction (MTP) head built into the model. Its targets are Qwen 3.8 Flash Next, Qwen 3.8 27B, and Ternary Bonsai 2 27B.
Speculative decoding is a way to speed up generation by validating draft tokens that the model produced first, all at once. The README states that it uses rejection sampling with residual correction for draft validation, and that the output is accurate at any temperature. The repository description cites 125 tok/s when using OpenCode on an M5 Max, and says a 27B model can run on a Mac with 16 GB of memory. Measurement conditions and record links are published for the figures. It also provides an OpenAI- and Anthropic-compatible local server, and the license is Apache-2.0.
The performance figures are results from the specific environment the repository presents; there is no ground for assuming they apply as-is to other hardware or usage patterns. The README also notes that equivalence checks for image requests have not been repeated yet. macOS 14 or later is required. Therefore the published performance claims should be read as records under those conditions rather than general comparison results, and the accuracy of image requests needs further confirmation.