6-11X speedup running Qwen3.6 on my RTX 4090, full repo.
Summary
Lightreason describes a local inference lab to speed up Qwen on a single RTX 4090. It reports speedups ranging from 6x to 11x across models and prompts, outlines setup and porting guidance, and credits CUDA Graphs, PARD2, and FreeToken as enabling techniques. The repository provides run instructions, benchmarks, and reproducibility notes for researchers and practitioners.