First direct benchmark of qwen3.6-27b vs qwen3.6-35b-a3b (both NVFP4) on a single dxg spark
35B is 5-6x faster
Prefill/decode:
27B➡️ reads 1.1k tok/s, types 13 tok/s
35B➡️ reads 6k tok/s, types 76 tok/s
A3B is MOE so this checks out
vLLM v0.24.0, single stream, 2K/8K/16K inputs
Real question is how good is each ones tool calling in hermes agent? Pending.