MTPLX boosts Apple Silicon MTP decode

What matters is not just the speedup, but that MTPLX claims to make temperature-correct speculative decoding practical on Macs, which is the real blocker most “fast decode” tools avoid.

• –The project’s core differentiation is exact probability-ratio acceptance at `temp > 0`, so it targets real coding/chat workflows instead of greedy-only demos
• –It leans on native MTP heads, which avoids the memory and maintenance cost of shipping a separate drafter model
• –The repo is positioned as a full local serving stack, not just a kernel trick: CLI, browser chat, terminal chat, metrics, and API compatibility are all part of the package
• –The benchmark headline is strong, but it appears tuned to a favorable prompt and thermal setup, so real-world gains will likely vary
• –Compatibility is still a real constraint because many MLX quants strip MTP heads, which limits immediate adoption