Qwen 3.6 Speed on 3090 vs. Apple Silicon

How is performance between local Qwen on Apple Silicon vs. a 3090?

The first layer that gets offloaded to RAM results in a fixed-cost memory copy back-and-forth to RAM. The cliff from 0->1 layers in RAM is expected.

Dashed blue and orange lines are the performance on my Mac Studio (M3 Ultra).

What does surprise me a bit? The stability of the results. Each point is 6 runs. The latency and performance were nearly identical across runs, especially for prompt processing.

Same graph but for 27B:

Takeaway? Performance is fucking great with the 35B A3B, and you can run the Q4 on a single 3090. It's even better in terms of tok/s output with MTP.

If you tried to 'cheat' and run it on a single GPU with offloading some RAM, the performance quickly dives below that of Apple Silicon.

the setup

• one-shot prompt
  • system prompt: Opus 4.6's system prompt (this necessitates processing a very long prompt)
  • 'write a minesweeper clone in one file.'
  • agents were instructed not to test so they don't waste time in tool calls.
• software
  • llama-server. q8 KV cache
  • Qwen 3.6 35BA3B - smallest Q4 variant from unsloth
  • harness: pi.dev
• model
  • qwen 3.6 Q4 from unsloth, not the MTP version

(I wrote this article; not an AI)