Hermes 4 70B
Release in the Hermes family · version Hermes-4-70B
Maintained by Nous Research15
A 70B-parameter hybrid reasoning model that Nous Research post-trained from Meta's Llama 3.1 70B base model, released in August 2025 with Hermes 4 405B and 14B. Nous trained it on a newly synthesized dataset of about 5 million reasoning and non-reasoning samples, and it can either answer directly or reason inside think tags first.156
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Paper: Hermes 4 technical report (arXiv 2508.18255) (external site: arxiv.org)
- Dataset hub: Evaluation samples (Hugging Face dataset, reasoning mode) (external site: huggingface.co)
- Repository: TorchTitan fork used for training (external site: github.com)
- Repository: Atropos environments (external site: github.com)
Availability and license
Overall availability
Weights download from Hugging Face without an access gate; an FP8 version is published separately. The model card lists Nous Portal, Chutes, Nebius, and Luminal as inference providers. Use is subject to Meta's Llama license terms.12
Availability is separate from permission: read the license before using or redistributing.
Meta Llama 3 Community License (card license field "llama3") (external site: raw.githubusercontent.com)110
The model card's license field is "llama3" and the repository has no LICENSE file. The base model, Llama 3.1 70B, is distributed under the Llama 3.1 Community License Agreement, which applies to derivative works: it requires including the agreement, displaying "Built with Llama", and beginning the names of distributed derived models with "Llama"; requires following Meta's acceptable use policy; requires licensees whose products had more than 700 million monthly active users in the month before the Llama 3.1 release to request a license from Meta; and ends the license for anyone who sues Meta or any other entity alleging that the Llama materials or their outputs infringe their IP. California law governs.141112
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
The weights are under a license that is not on the rubric's OSI-approved list. Read its terms before use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | BF16 safetensors in an ungated Hugging Face repository; an FP8 repository is also published.12 |
| Inference codeIs code for running the model published? | Public | The model card gives a Transformers example and says vLLM and SGLang include tool-call parsers for the Hermes format; the repository's config declares the standard LlamaForCausalLM architecture.13 |
| Training codeIs the code used to train the model published? | Partial | The technical report links the specific commit of Nous's TorchTitan fork (BSD-3-Clause) used for Hermes 4 training and says all Atropos environments used for rejection sampling are open source. This covers Nous's post-training only; Meta has not published the code used to pretrain the Llama 3.1 base.589 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The report describes the roughly 5-million-sample, 19-billion-token post-training dataset: DataForge synthetic data seeded from DCLM and FineWeb, rejection-sampled trajectories from Atropos environments, and a retained portion of the public Hermes 3 dataset. The Hermes 4 dataset itself is not linked as a release, and the base model's pretraining data is not public.5 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | For Nous's post-training, the report gives the parallelism, 56B training tokens, learning rate 1e-5, 12,864 B200 hours, 9,000 steps with 300 warmup steps, a global batch of 384 samples at 16,384-token context, packing, and loss masking. The base model's pretraining recipe is Meta's and is not part of this release.5 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Public | Nous released the samples generated during evaluation as Hugging Face datasets (reasoning and non-reasoning modes), and the report describes the harness (lighteval, including a published Nous branch, EQBench, and Atropos) and sampling settings.57 |
What it is useful for
Nous presents it as a general assistant for math, code, STEM, and creative tasks with function calling and tool use and schema-following JSON output, and describes it as trained to be easier to steer, with reduced refusal rates.1
Run and use notes
- The model card recommends temperature 0.6, top_p 0.95, and top_k 20 and uses the Llama 3 chat format, with reasoning switched on through a thinking flag in the chat template or a system prompt.1
Organization context
Provenance and derivatives
Fine-tuned by Nous Research from Meta's Llama 3.1 70B base model. Nous Research's contribution is the synthesized post-training data and the supervised fine-tuning run; the base model was pretrained by Meta.1512
- Derived from: Llama 3.1 70B (meta-llama/Llama-3.1-70B) — Meta base model; see the Llama family record. The card's base_model field gives the earlier repository name meta-llama/Meta-Llama-3.1-70B, which now redirects here.
Other releases in the Hermes family
- Hermes 4.3 36BModel-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Eligible · basis: U.S. headquarters
This fine-tune was made and published by Nous Research, Inc., a Delaware corporation with U.S. addresses (see the nous-research record). The Llama 3.1 70B base model is Meta's; this record's eligibility rests on Nous Research's post-training and does not re-attribute the base model.1135
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.