> For the complete documentation index, see [llms.txt](https://brindha.gitbook.io/mylearning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://brindha.gitbook.io/mylearning/tools/hermes.md). # Hermes **Hermes** is a family of fine-tuned large language models developed by **Nous Research**, an open-source AI research collective. Unlike most AI models on this list, Hermes is not built from scratch but is instead a series of carefully fine-tuned versions of existing base models — most notably **Llama** and **Mistral** — optimized for instruction following, reasoning, and agentic tasks. *** #### What It Is Hermes is a community-driven, open-source model series that takes powerful base models and fine-tunes them using high-quality curated datasets. Nous Research focuses on improving **instruction following, structured output, tool use, and reasoning** beyond what the original base models offer out of the box. Hermes has become one of the most respected and widely used fine-tuned model families in the open-source AI community. *** #### Key Capabilities * **Instruction following** — highly refined ability to follow complex, multi-step instructions * **Structured output generation** — reliable JSON, XML, and other formatted outputs * **Function calling and tool use** — strong support for agentic workflows and API integrations * **Reasoning and analysis** — improved logical and mathematical reasoning over base models * **Code generation and debugging** — capable across multiple programming languages * **Roleplay and creative writing** — one of the strongest open models for creative and character-based tasks * **Long context understanding** — handling large documents and extended conversations * **Conversational AI** — natural, nuanced dialogue with strong personality and depth *** #### Versions and Models | Version | Notes | | -------------- | ----------------------------------------------------------------------------------- | | Hermes 1 | Early fine-tune of GPT-4 class models; established Nous Research reputation | | Hermes 2 | Fine-tuned on Llama 2; significant improvement in instruction following | | Hermes 2 Pro | Enhanced function calling and structured output capabilities | | Hermes 2 Theta | Merged model combining Hermes and Dolphin weights | | Hermes 3 | Latest generation; fine-tuned on Llama 3; major leap in reasoning and agentic tasks | | Hermes 3 405B | Largest variant based on Llama 3.1 405B; among the most powerful open models | *** #### Access Options * **Hugging Face** — all model weights freely available for download * **Ollama** — easy local deployment of Hermes models * **LM Studio** — desktop application for running Hermes locally * **Together AI and other API platforms** — hosted inference for Hermes models * **OpenRouter** — API aggregator offering access to Hermes variants * **Self-hosting** — full weights available for private deployment on own hardware *** #### Strengths * **Exceptional instruction following** — widely regarded as among the best open models for following complex instructions precisely * **Structured output reliability** — highly consistent JSON and formatted output generation, critical for agentic applications * **Strong function calling** — one of the best open-source models for tool use and API integrations * **Creative and roleplay capability** — particularly strong for character-based writing and creative tasks * **Fully open-source** — weights freely available with permissive licensing * **Community-driven** — developed by a dedicated research collective with strong community involvement * **No commercial restrictions** — can be used freely for commercial applications * **Regularly updated** — Nous Research consistently releases new versions tracking the latest base models *** #### Limitations * **Not a base model** — dependent on upstream base models like Llama; improvements are bounded by the base model's capabilities * **No built-in interface** — requires technical knowledge or third-party tools to deploy and use * **No real-time web search** — operates within a knowledge cutoff with no live data access * **Hardware requirements** — larger variants like 405B require significant compute resources * **Smaller organization** — Nous Research is a small collective with limited resources compared to major AI labs * **Less name recognition** — not widely known outside of the open-source AI developer community *** #### Use Cases Hermes is widely used for: * **Agentic AI applications** — building autonomous agents that use tools and APIs reliably * **Structured data extraction** — parsing and generating reliable JSON outputs from unstructured text * **Chatbot and assistant development** — creating custom AI assistants with strong instruction following * **Creative writing and roleplay** — character-based storytelling and narrative generation * **Local AI deployment** — running a capable assistant privately on personal hardware * **Research and experimentation** — studying fine-tuning techniques and model behavior * **Enterprise automation** — building reliable pipelines that require consistent formatted outputs * **Function calling workflows** — integrating AI into software systems via tool use *** #### Nous Research and the Open-Source Community Nous Research occupies a unique position in the AI landscape: * Operates as a **volunteer-driven research collective** rather than a traditional company * Focuses entirely on **advancing open-source AI** through fine-tuning and dataset curation * Has produced some of the **most widely downloaded models** on Hugging Face * Shares research findings and datasets openly with the community * Collaborates with other open-source projects to advance the broader ecosystem * Receives support from community donations and partnerships rather than venture capital funding *** #### Why Choose Hermes? Hermes is an excellent choice if you value: * **Reliable instruction following** — precise adherence to complex and detailed prompts * **Agentic capabilities** — building AI systems that use tools and APIs effectively * **Structured output generation** — consistent and well-formed formatted responses * **Creative flexibility** — strong performance on roleplay and character-driven tasks * **Open-source freedom** — full access to weights with no usage restrictions * **Local deployment** — running a high-quality model privately on your own hardware * **Community-backed development** — trusting a model refined by a dedicated research community *** Hermes represents the best of what the **open-source AI fine-tuning community** can achieve — taking powerful foundation models and refining them into highly capable, reliable, and versatile assistants that serve developers, researchers, and creative users alike, all without the restrictions of proprietary systems.