Hermes Tutorial
In one sentence: Hermes is Nous Research's open-source AI project, spanning the hybrid-reasoning Hermes 4 model family and the Hermes Agent that learns and improves over time.
What is Hermes?
Hermes is developed and open-sourced by Nous Research, with two main lines:
- Hermes 4 models: An open-source hybrid-reasoning LLM family released in 2025, fusing structured multi-step reasoning with instruction-following. It leads on math, coding, and logical reasoning.
- Hermes Agent: An open-source, self-improving AI agent that generates skills from experience, optimizes continuously, accumulates knowledge, and builds a dynamic model of your preferences across sessions.
Hermes 4 Model Versions
| Version | Base Model | Highlights |
|---|---|---|
| Hermes-4-14B | Qwen 3 14B | Lightweight, ideal for local deployment |
| Hermes-4-70B | Llama 3.1 70B | 131k context, well-balanced choice |
| Hermes-4-405B | Flagship | 405B parameters, strongest capability |
| Hermes 4.3 36B | ByteDance Seed 36B | First Hermes model trained in a decentralized way via Psyche |
Highlights
- Hybrid reasoning mode: Explicit
<think>blocks reveal the reasoning process; switch between fast responses and deep reasoning. - Strong reasoning: Notable gains in math, code, STEM, and logic, while also handling creative writing.
- Structured output: Generates valid JSON matching a given schema and auto-repairs malformed objects.
- Top of RefusalBench: Outperforms major closed- and open-source models on uncensored helpfulness while respecting user values.
- Tool use: Native ChatML format and function calling, compatible with vLLM, SGLang, and other inference frameworks.
Quick Start
bash
# Try it online (no install)
# Visit https://chat.nousresearch.com to chat
# Hermes Agent one-liner install (only requires Git)
git clone https://github.com/NousResearch/hermes-agent && cd hermes-agent && ./start.shUse Cases
- Research & experiments: Open weights can be deployed locally and fine-tuned — great for model research.
- Programming & reasoning: Excellent at math, code generation, and complex logical analysis.
- Agent development: Hermes Agent suits autonomous workflows with learning and memory.
Note: Nous Research recommends the Hermes 4 models for conversation and reasoning. For agent workflows that depend on frequent tool calls, pair them with models better at tool use, such as DeepSeek or Claude.
Hermes open-source repos: github.com/NousResearch