Hermes 4 Technical Report (arXiv:2508.18255), Table 3 — 'Final downstream task evaluations for Hermes 4 405B and comparable open-weight models'

Read https://arxiv.org/pdf/2508.18255 for the benchmark figures on models/hermes-4-405b. As recorded by the reader: Values are the Hermes 4 405B reasoning-mode column (R) of Table 3; the card's own HF page shows benchmarks only as an image and points to this report ('Full tables, settings, and comparisons are in the technical report'). Table cell format is score(non-reasoning): MATH-500 96.2(73.8), AIME'24 81.9(11.4), AIME'25 78.1(10.6), GPQA Diamond 70.6(39.4), LCBv6 Aug2024+ 61.4(28.1), MMLU 87.2(73.6), MMLU-Pro 80.6(58.3), IFEval (Loose) 81.5(84.9), Arena-Hard v1 93.7(53.5). Source label is 'LCBv6 Aug2024+' which maps to LiveCodeBench v6; IFEval is the 'Loose' accuracy variant. Evaluated with temperature 0.6 / top-p 0.95 on SGLang 0.4.9.post3, B200 GPUs.