Hermes 3
Open-weight instruct model based on Llama-3.1 405B
Hermes 3 is a generalist language model developed by Nous Research, released in August 2024. It is a full parameter fine-tune of the Llama-3.1 405B foundation model, designed for advanced agentic capabilities, roleplaying, and tool use. The model offers strong reasoning skills and structured output control for users.
- Full parameter finetune of Llama-3.1 405B
- Uses ChatML prompt format for structured dialogue
- Supports advanced function calling and JSON modes
- FP8 quantization reduces VRAM need to ~430GB
- Released by Nous Research in August 2024
What it is
Hermes 3 is a large language model created by Nous Research. It serves as the flagship entry in the Hermes series. The 405 billion parameter version is a full parameter fine-tune of Meta’s Llama-3.1 405B foundation model. Released in August 2024, it focuses on aligning the model to user needs through improved steering capabilities. It supports general assistant tasks, complex reasoning, and creative roleplaying. The model is designed to give end users significant control over its behaviour and output format.
How it came about
This model expands upon the capabilities established in Hermes 2. Nous Research aimed to improve reliability in function calling and structured outputs. The team trained the model using compute resources provided by LambdaLabs via their 1-Click Cluster service. The technical report was published on arXiv in August 2024 by Ryan Teknium, Jeffrey Quesnelle, and Chen Guang. It represents the first full parameter fine-tune since the release of the underlying Llama-3.1 405B base model.
Inside the model
The architecture is a decoder-only transformer derived from Llama-3.1. It uses the ChatML prompt format, which structures multi-turn dialogue with special tokens for roles like system, user, and assistant. This format allows for detailed system prompts that guide rules and stylistic choices. The model was trained to handle function calling by parsing JSON schemas within XML tags. It also supports a specific mode for structured JSON outputs, adhering strictly to provided Pydantic schemas. Inference in full precision requires over 800GB of VRAM, so quantized versions are available.
What it can do
Hermes 3 demonstrates strong performance in general capabilities, often matching or exceeding Llama-3.1 Instruct models. It excels in multi-turn conversation coherence and long context handling. The model shows improved skills in code generation and reliable function calling. It can act as an agent by selecting tools and interpreting their results. Benchmark results indicate it is competitive among open weight models. Its design supports advanced roleplaying scenarios where the model adopts specific personas or emotional states.
How to use it
The model weights are openly available for download. Users can access the 405B variant through Hugging Face. Due to its size, running the full precision model requires substantial hardware. Nous Research provides FP8 quantized versions that fit into approximately 430GB of VRAM, compatible with the VLLM inference engine. Alternative loading methods include 4-bit or 8-bit quantization using bitsandbytes, though this may be slower. The prompt format ensures compatibility with OpenAI-style API endpoints.
Questions people ask about Hermes 3
What is the base model for Hermes 3?
Hermes 3 is built on top of the Llama-3.1 405B foundation model. It is a full parameter fine-tune, meaning all weights were adjusted during training rather than just a small subset.
How much memory is needed to run Hermes 3?
The full precision model requires over 800GB of VRAM. However, an FP8 quantized version is available that fits into approximately 430GB of VRAM, making it more accessible for high-end GPU clusters.
Does Hermes 3 support tool use?
Yes, it has strong function calling capabilities. It can parse JSON schemas within XML tags to select and execute tools, then interpret the results to provide natural language responses.
Good reading and sources
There is no English Wikipedia article on Hermes 3 yet, so the facts on this page come from the sources listed above: huggingface.co/NousResearch/Hermes-3-Llama-3.1-405B, arxiv.org/abs/2408.11857. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.