AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuned / chat models

Hermes-4-Llama-3.1-405B

Nous Research
open weights / Overall score: 3.1

Nous Research's hybrid-mode reasoning instruct line, one post-training recipe shipped on four independent base models: ByteDance's Seed-OSS-36B, Qwen3-14B, and Meta's Llama 3.1 at 70B and 405B. A single checkpoint toggles between reasoning and non-reasoning modes.

The four base models carry different licenses and a user can substitute freely between them, so the entry follows the 36B build rather than the most restrictive base.

Openness

3 high confidence
3.0
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)

The Hermes 4 recipe ships on four base models, and the license follows whichever a user picks: Apache-2.0 on the Seed-OSS-36B and Qwen3-14B builds, Meta's more restrictive terms on the Llama-3.1 builds; since switching to the Apache build is free, the score follows it. Its post-training data is proprietary and only inference code ships, which keeps the family at open weights rather than higher.

  • https://huggingface.co/api/models/NousResearch/Hermes-4.3-36B recorded 2026-09-29

    The Hugging Face API record for the governing release lists "license:apache-2.0", "base_model:ByteDance-Seed/Seed-OSS-36B-Base" and "gated":false, so the weights download without an access gate.

  • https://huggingface.co/NousResearch/Hermes-4-405B-FP8 recorded 2026-09-29

    The Llama-based build of the same recipe is tagged "license:llama3" with "base_model:meta-llama/Llama-3.1-405B"; it is the build whose terms do not govern, because a user wanting Apache terms takes the Seed-OSS or Qwen build.

  • https://huggingface.co/NousResearch/Hermes-4.3-36B recorded 2026-09-29

    The governing release card carries the repo tag "license:apache-2.0" and the base model tag "base_model:ByteDance-Seed/Seed-OSS-36B-Base"; searched the body for post-training mixture, training code and recipe release: only a benchmark table and prompt format are published, with no post-training data or training code.

Adoption

1 high confidence
1.0

Adoption is measured as Hugging Face downloads summed across the four declared Hermes checkpoints. Traffic is spread across the Hermes 3 and Hermes 4 builds, with no single checkpoint drawing most of it.

Capability

4 medium confidence
4.0

The strongest Hermes 4 checkpoint is near-frontier on math and knowledge for an open instruct model, scoring 61.4 on LiveCodeBench v6. Its coding still trails the 2026 frontier coders such as DeepSeek-V4, Qwen3.6 and Kimi, which keeps it a step below the top tier.

  • 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

Verified 2026-09-29, except some scores, still awaiting re-confirmation