FINAL-Bench/Darwin-180B-RSI model card

Fetched 2026-10-18. HF API createdAt 2026-09-28T03:15:04Z, lastModified 2026-10-05T23:32:23Z. Safetensors parameter total reported by the HF API is 179,999,981,424 BF16 plus 35 I64, i.e. 180.000B. The card says the checkpoint is a full-parameter self-improvement of Qwen/Qwen3.8-Flash-Next (125B body, 6B activated, plus the n-gram table and MTP heads that bring the file to 180B). Benchmark numbers live in .eval_results/*.yaml on this repo and are the publisher own runs: GPQA Diamond 94.44 (all 198 items, majority vote over up to 16 samples, temperature 1.0, top_p 0.95, top_k 20, thinking budget 131072, bf16), AIME 2026 100.0 (maj@16; the same file also records mean accuracy 98.75), MMLU-Pro 88.12 (single sample, all 12032 items, same sampling). All three are vendor claims. The card also claims first place on MDPBench, IFStruct, MMMU-Pro, LEXam and LEXam-hard, which are outside the benchmarks this repo ranks on. The LICENSE file in the repo is the Qwen Community License 1.0, inherited from the base.