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.