Benchmark Same 32 models on M2 Pro, M3 Ultra M4 Pro and M5 Max
retained text
I had four Macs free overnight and ran the same benchmark on all of them: same build, same models, same settings. Decode speed, tok/s (4-bit, no speculative decoding): |Model|M2 Pro 32GB|M4 Pro 48GB|M5 Max 36GB|M3 Ultra 256GB| |:-|:-|:-|:-|:-| |Qwen3.5 4B|63|79|133|159| |gpt-oss 20B|50|74|114|132| |Gemma 4 26B (MoE)|53|73|115|118| |Qwen3-Coder 30B (MoE)|–|89|136|100| |Qwen3.8 27B (dense, 3.5bpw)|12|16|28|37| One 1024x1024 image with Qwen-Image 2.1: M2 Pro 451 s, M4 Pro 264 s, M5 Max 38 s, M3 Ultra 83 s. What surprised me: * On MoE models the 36 GB M5 Max matches or beats the M3 Ultra. Only 3-4B params are active per token, so the newer GPU matters more than bandwidth. * On the dense 27B the Ultra still wins (37 vs 28). That one is bandwidth-bound. * Image generation is the biggest jump: the M5 Max is \~7x faster than the M4 Pro and \~2x faster than the Ultra. * M2 Pro -> M4 Pro is only \~1.3-1.5x on text. Method: prompt 512 tokens, generate 128, 1 warmup + 5 runs, median. One run per cell, so treat small gaps as noise. MLX-based engine (disclosure: I work on Rapid-MLX, which is what I used); other runtimes will give different absolute numbers, the Mac-to-Mac ratios are the interesting part. "–" = doesn't fit or not run.
document fetched 2026-10-31 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xjp58z/benchmark_same_32_models_on_m2_pro_m3_ultra_m4