MTP test on Mac Studio: +92% on dense model, -19% on MoE

A developer benchmarked MTP speculative decoding in oMLX on a Mac Studio with M2 Ultra: code generation on dense Qwen3.8-27B sped up 92%, while prose on the Ornith-1.5-35B-A3B MoE model slowed 19%. EAGLE-4 itself does not run on Apple Silicon, requiring Linux and four Nvidia GPUs.

• Qwen3.8-27B: code without thinking — 33.6 vs 64.3 tok/s, 1.92x speedup
• Ornith-1.5-35B-A3B: prose without thinking — 133.1 vs 107.6 tok/s, 0.81x slowdown
• EAGLE-4 needs Linux, 4 Nvidia GPUs and CUDA 13; no Metal or MLX support
• At temperature 0, MTP changed output text in 7 of 8 test cases