Three Qwen Models, One DGX Spark, Five Real Agents
I replayed 76 real requests from my self-hosted Hermes agents against Qwen3.8-Flash-Next, Qwen3.8-27B with SGLang + DFlash2 and Qwen3.6-35B-A3B with vLLM + MTP. Here are the numbers on TTFT, decode tok/s, concurrency and a blind quality review.
16 min
AI
LLM
Self-Hosted
Homelab
DevOps
