The DeepSeek-V4.1-Flash-1017 Q2 GGUF model runs natively on a single RTX 3090 with 192 GB RAM under Windows (no WSL) and delivers an average throughput of 41 tokens/s across multiple prompts. It matches the official DeepSeek API’s next‑token predictions 90.2 % of the time on general text and 92.4 % on code/math, staying within a single‑token deviation. The implementation supports the full 1 M‑token context window and includes a lightweight inference interface.