VIDRAFT's Darwin-180B-RSI, a 180‑billion‑parameter open‑weight language model hosted on Hugging Face as FINAL‑Bench/Darwin‑180B‑SI, achieved the top rank on AIhoster’s independent open‑model benchmark board for both MMLU‑Pro and GPQA Diamond as of the 2026‑10‑19 update. On MMLU‑Pro it scored 88.1% accuracy across the full 12,032‑item test set, surpassing larger models such as DeepSeek‑V4‑Pro (1.6 T parameters, 87.5%) and Kimi‑K2.5 (1 T parameters). On GPQA Diamond it attained 94.4% accuracy on all 198 items, outperforming Kimi‑K3 (2.8 T parameters, 93.5%) and the VIDRAFT sibling Darwin‑397B‑ZTC (404 B parameters, 93.4%) which used greedy decoding. Both evaluations employed an extended thinking budget of up to 131,072 tokens; MMLU‑Pro relied on a single sample (temperature = 1.0, top_p = 0.95, top_k = 20), whereas GPQA Diamond applied majority voting over up to 16 samples under the same sampling configuration. The model runs in bf16 precision, indicating strong parameter‑efficient reasoning likely derived from architectural choices such as mixture‑of‑experts or distillation, though specific details remain undisclosed. The weights are accessible via the Hugging Face CLI command `huggingface-cli download FINAL-Bench/Darwin-180B-RSI`.