Running Qwen3.8 at 255 tok/s on a single 5090

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.