Qwen3.5-122B-A10B hits 30 tps on M1 Ultra

This reads less like a one-off curiosity and more like evidence that Qwen3.5’s 10B-active MoE design can make very large models practical on consumer Macs. The real story is the decode speed: 30.5 tps is not “desktop toy” territory for a 122B model.

• –The 396 tps prompt speed suggests prefill is very healthy, even with a 22K-token context, so long-context workflows may be more usable than people expect.
• –The 30.5 tps generation rate is the number that matters for interactive use, and it’s respectable for a local 122B-class model on Apple Silicon.
• –llama.cpp via Unsloth Studio is a solid baseline, but MLX is the obvious next comparison point for Mac owners chasing more throughput.
• –This is still a single user report, so quant choice, sampler settings, and context length could move the numbers a lot.
• –The broader takeaway: Qwen3.5-122B-A10B is no longer automatically “too big for Mac,” but it remains highly setup-sensitive.