Lightreason Achieves Up to 11X Speedup for Qwen Models on RTX 4090
November 13, 2026
The Lightreason inference lab uses CUDA graphs and compact KV caching to boost Qwen2.5-3B-Instruct from 63.0 to 139.6 tok/s. For Qwen3.6-35B-A3B, the PARD2 runner reached up to 724.4 tok/s on specific Fibonacci slices using system RAM for expert weights.
HOW THIS AFFECTS YOU
●
builderYou can get much higher throughput on consumer hardware by using these specific optimization techniques.