The developer ported the Splash inference framework (originally M3+ only) to Apple's 2021 M1 Max, enabling Qwen3.8‑27B to run at ~39 tokens per second using custom Metal kernels. A pre‑built release eliminates the need for Xcode or compilation, installable via a single curl‑pipe command and sharing the model cache with the official Splash build. The same approach also supports the 35B MoE variant, achieving ~94.7 tok/s average on npanj's benchmark and peak rates near 450 tok/s.