Hermes-4-Llama-3.1-405B

weights

open(Apache-2.0)

base

Seed-OSS-36B-Base

post-training-data

closed(proprietary ~5M-sample corpus)

code

partial(inference only)

license

Apache-2.0(OSI)

Nous ships one Hermes 4 post-training recipe on four independent base models, and the base determines the license the weights carry: Seed-OSS-36B and Qwen3-14B are Apache-2.0, while the Llama-3.1 405B and 70B builds inherit Meta's Llama 3.1 Community terms (a 700M monthly-active-user commercial cap, an acceptable-use policy, naming requirements). Scored 3 on the Apache builds rather than 2 on the Llama ones, because resolving to the most restrictive SKU is a rule about variants you cannot substitute away from - Qwen 2.5's restricted 72B caps Qwen because the 7B is no substitute for the flagship - and here the base is a free choice. Anyone wanting Apache-licensed Hermes 4 takes the 36B or 14B build and gets it, so the family genuinely offers open weights. Not 4 or 5: the post-training corpus is proprietary and only inference code ships, so the recipe is not reproducible. The governing release, NousResearch/Hermes-4.3-36B, is Apache-2.0, ungated and built on ByteDance-Seed/Seed-OSS-36B-Base, which is the fact the 3 rests on; neither it nor the 405B Llama-based build releases the post-training corpus.

• https://huggingface.co/NousResearch/Hermes-4.3-36B recorded 2026-08-13

  the governing release - `license:apache-2.0` in the repo tags, `"gated":false`, and `base_model:ByteDance-Seed/Seed-OSS-36B-Base`; the card publishes prompt format and benchmarks only, with no post-training mixture and no training code

• https://huggingface.co/NousResearch/Hermes-4-405B-FP8 recorded 2026-08-13

  the Llama-based build of the same recipe - `license:llama3`, base Meta-Llama/Llama-3.1-405B; the SKU whose terms do NOT govern, because a user wanting Apache takes the Seed-OSS or Qwen build

The four declared SKUs sum to 34,445 Hugging Face downloads in the trailing 30 days (Hermes-4.3-36B 18,158; Hermes-3-Llama-3.1-70B 10,217; Hermes-4-14B 5,401; Hermes-3-Llama-3.1-405B 669), which is level 2 - early-adopter scale. The 36B is the most-downloaded SKU, its small size and permissive license driving the uptake.

• https://huggingface.co/api/models/NousResearch/Hermes-4.3-36B recorded 2026-08-13

  18,158 downloads in the trailing 30 days for NousResearch/Hermes-4.3-36B

• https://huggingface.co/api/models/NousResearch/Hermes-4-14B recorded 2026-08-13

  5,401 downloads in the trailing 30 days for NousResearch/Hermes-4-14B

The strongest Hermes 4 SKU is near-frontier on math and knowledge for an open instruct model, though its coding trails the 2026 frontier coders, so not a category-defining 5 (those are DeepSeek-V4, Qwen3.6 and Kimi class). The technical report's coding row is LiveCodeBench v6 at 61.4. The OpenRouter model page carries no benchmark figure, so it evidences distribution rather than capability.

• https://arxiv.org/pdf/2508.18255 recorded 2026-08-13

  Hermes 4 technical report, 405B reasoning column - MATH-500 96.2, AIME'24 81.9, AIME'25 78.1, GPQA Diamond 70.6, BBH 86.3, LCBv6 (Aug2024+) 61.4, MMLU 87.2, MMLU-Pro 80.6, with non-reasoning figures in parentheses beside each

• https://openrouter.ai/nousresearch/hermes-4-405b recorded 2026-08-13

  Hermes 4 405B listed and routable on OpenRouter; the page renders no benchmark figures to a plain fetch, so it evidences distribution rather than capability