Post #233880 684 Oct 13, 2026, 22:43 UTC Running Qwen3.8 at 255 tok/s on a single 5090Article, Comments mesh.applied mesh.applied | Task-specialized open models for your inference needs Task-specialized open models built for your inference needs: DeepSeek and Qwen, taken through post-training, ternary and NVFP4 quantization and custom kernels, so they run fast on local RTX hardware. 🔥 9 🌭 1