Nous: Hermes 3 405B Instruct

Hermes 3 405B is the flagship of Nous Research's Hermes 3 series, published on Hugging Face as NousResearch/Hermes-3-Llama-3.1-405B. The model card calls it a "frontier level, full parameter finetune" of Meta's Llama-3.1-405B, aimed at aligning the model to the user. Its technical report, "Hermes 3 Technical Report" (arXiv 2408.11857), was published on August 15, 2024 by Ryan Teknium, Jeffrey Quesnelle and Chen Guang. The repository remains online as of October 29, 2026. This entry covers the Hugging Face release of the model; it is not an API service.

Key Features

Base and size: fine-tuned from meta-llama/Llama-3.1-405B, with about 406B parameters in BF16 Safetensors.
Capabilities claimed by Nous: function calling and structured output (JSON mode), agentic abilities, roleplaying, reasoning, multi-turn conversation and code generation. It uses the ChatML prompt format with a system prompt for steerability.
Competitiveness: the card says Hermes 3 is competitive with, and in some areas possibly superior to, Llama-3.1 Instruct models, with strengths and weaknesses varying. The detailed comparison is in a chart on the card.
License: the Llama 3 license, inherited from the base model.
Compute: the card credits LambdaLabs and its 1-Click Cluster service.

Use Cases

Agent and function-calling research with a large open model.
Roleplay, long multi-turn dialogue and steerable system-prompt use.
A base for further fine-tuning or distillation experiments.

Limitation: the model is very large. The card says FP16 loading needs over 800GB of VRAM, while the pre-quantized FP8 build (NousResearch/Hermes-3-Llama-3.1-405B-FP8) needs about 430GB. The card states no context length, and no Inference Provider hosts it on Hugging Face.

Pricing and availability

The weights are free to download under the Llama 3 license terms. The card lists no hosted price; this entry quotes none. Usage examples are given for Transformers, vLLM, SGLang and Docker Model Runner.

Getting Started

Check the Llama license terms, then download the BF16 or FP8 weights.
Serve with vLLM or SGLang across multiple GPUs, using the FP8 build if memory is limited.
Format prompts with ChatML and a system message.

FAQ

What is Hermes 3 405B based on?

Meta's Llama-3.1-405B, per the model card.

How much memory does it need?

Over 800GB of VRAM at FP16, and about 430GB for the FP8 build, according to the card.

Does it support function calling?

Yes, the card lists function calling and JSON-mode structured output.

Is there a smaller Hermes?

The catalog has Hermes 2 Pro Llama-3 8B, an earlier and smaller Nous model.

Alternatives

Llama 3.1 405B Instruct: Meta's own instruct model on the same base.
Llama 3.1 70B Instruct: a far smaller Meta option.
Hermes 2 Pro Llama-3 8B: a lightweight earlier Hermes.

Conclusion

Hermes 3 405B is a steerable, tool-capable open fine-tune that needs serious hardware. For a lighter option, try a smaller Llama or Hermes model. See more in the models category.