Hermes 4 70B

Hermes 4 70B is Nous Research's 70B-parameter hybrid-mode reasoning model, a full-parameter fine-tune of Meta Llama-3.1-70B trained on ~5M samples / ~60B tokens of blended reasoning and non-reasoning data. It can either respond directly or deliberate inside explicit <think>…</think> traces, with a second training stage that caps reasoning length at ~30K tokens to stop overlong "thinking loops". It supports structured outputs (JSON mode, schema adherence), function calling and tool use, and is aligned for high steerability with broadly reduced refusal rates (SOTA on Nous's own RefusalBench). Weights are freely downloadable on Hugging Face in BF16, FP8 and GGUF variants; the model ships under the Llama 3.1 Community License, so check Meta's use restrictions before commercial deployment.

Provider-reported (Nous eval dataset on HF, NousResearch/eval-Hermes-4-70B-reasoning): AIME 24 73.5% · AIME 25 67.4% · MATH-500 95.6% · MMLU 88.4% · MMLU-Pro 80.7% · GPQA-Diamond 66.1% · ArenaHard 90.1% · Creative Writing v3 77.5%.

Independent: Artificial Analysis lists Intelligence Index 12.6 / Coding 9.2 / Math 11.3 (Apr 2026, measured on live inference endpoints). Press coverage (VentureBeat) frames it as competitive with ChatGPT-class models; no independent academic replication of the full table was found.