A new open‑source repo shows DeepSeek‑V4.1‑Flash‑1017 running in FP8 on 4x RTX PRO 6000 GPUs. The model achieves 166‑211 tokens per second, 722 tokens per second at 8 streams, and supports 1M context, vision, and tools. This demonstrates high‑throughput inference for developers.
Opening Kapyn…