NousResearch Hermes Complete Guide — Hermes 4.3 36B, Function Calling & Hermes Agent [2026]

NousResearch Hermes is a series of open-source LLM fine-tunes built on the philosophy of "user-aligned, minimally filtered, and highly steerable." From Hermes 1 (2023) through Hermes 4.3 (late 2025), the series has consistently led open-source LLMs in Function Calling reliability and agentic use cases. In March 2026, NousResearch also released "Hermes Agent," an open-source agent framework that has already surpassed 40,000 GitHub stars.

Hermes 4.3 is the first Hermes fine-tune based on a non-Meta model — ByteDance Seed 36B. It delivers 70B-class performance in a 36B Dense architecture with a remarkable 512K token context window. Training was conducted on the Psyche distributed network (Solana-based), and the model achieves the highest score of any model on RefusalBench.

RefusalBench measures how often an LLM unnecessarily refuses legitimate user requests. GPT-4o and Claude both score around 17%, while Hermes 4 reaches 57% and Hermes 4.3 exceeds that further. This is not about removing safety guardrails — it reflects NousResearch's philosophy of responding to legitimate requests without over-filtering. For AI agents and business automation, where over-refusal breaks workflows, Hermes's high RefusalBench score is a meaningful practical advantage.

Since Hermes 2 Pro, the series has used dedicated tokens (`<tools>`, `<tool_call>`, `<tool_response>`) enabling streaming-compatible, highly reliable Function Calling. It is widely regarded as the most dependable open-source LLM for FC implementations and also supports structured JSON output conforming to JSON Schema. Below is an example in ChatML prompt format:

<|im_start|>system
You are a helpful assistant. You have access to the following tools:
<tools>
[{"name": "get_weather", "description": "Get current weather", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}}]
</tools>
<|im_end|>
<|im_start|>user
What is the weather in Tokyo?
<|im_end|>
<|im_start|>assistant
<tool_call>
{"name": "get_weather", "arguments": {"location": "Tokyo"}}
</tool_call>
<|im_end|>
<|im_start|>tool
<tool_response>
{"temperature": 18, "condition": "Partly cloudy"}
</tool_response>
<|im_end|>

Hermes Agent is an open-source AI agent framework released in March 2026, already garnering 40,000+ GitHub stars. Built around the concept of a "growing agent," it automatically generates and memorizes new skills after completing tasks. Key features include persistent cross-session memory, natural-language Cron scheduling, multi-platform messaging (Telegram, Discord, Slack, LINE, WhatsApp, and more), official Ollama integration, and sub-agent support for parallel task execution.

Choose Hermes Agent if auto skill generation and broad messaging integrations are priorities. Choose OpenClaw if you need a mature MCP ecosystem or a proven production workflow.

Hermes 3 is available in Ollama's official library, enabling one-command setup. Hermes 4.3 requires a community-provided Modelfile.

Hermes 3 3B at Q4 quantization can run on CPU only, though response speed is significantly reduced.

1. AI Agent Backend LLM — Hermes's reliable Function Calling makes it the go-to choice for autonomous agent core engines. 2. Chatbot Development — Low filtering enables natural, useful responses in business chatbot deployments. 3. Creative Writing — Hermes 3 is particularly well-regarded for creative text generation tasks. 4. Structured Data Extraction — JSON Schema-compliant output makes it easy to integrate into data pipelines. 5. Multi-Step Reasoning — Hermes 4's hybrid reasoning mode handles complex logical and analytical tasks.

Hermes models are primarily trained on English data. For Japanese natural language generation, Qwen 3.5 or Gemma 4 are recommended instead. However, since Function Calling and JSON output are largely language-agnostic, Hermes remains fully usable for tool integration and data extraction workflows — even when user inputs are in Japanese. Structuring prompts so that tool calls operate in English while accepting Japanese inputs is a practical approach.

DeepHermes 3 was the first model to allow reasoning mode to be toggled via the system prompt. When enabled, it generates extended reasoning chains enclosed in `<think>...</think>` tags, significantly improving scores on math and logic tasks. Disabling it for conversational tasks keeps latency low. This dual-mode capability means a single model can cover both routine and high-precision tasks — a practical advantage for production deployments.

Hermes 4.3 was trained on the Psyche distributed network. Using the DisTrO optimizer, training was distributed across multiple data centers over the internet, with consensus on compute contributions and rewards secured by the Solana blockchain. This represents a significant innovation: large-scale model training without reliance on any single centralized compute provider, demonstrating that decentralized AI training is viable at frontier scale.

Q: What is the difference between Hermes and Llama? A: Llama is a foundation model released by Meta. Hermes is a fine-tune of Llama (and other base models) using NousResearch's proprietary high-quality datasets, optimized for Function Calling and agentic tasks. Q: Can Hermes be used commercially? A: It depends on the version. Hermes 3 follows the Meta Llama License; Hermes 4.3 follows the ByteDance Seed License. Review each license before commercial deployment. Q: Is Function Calling better than other open-source models? A: Yes — Hermes is considered the most reliable open-source LLM for Function Calling, thanks to its dedicated token system introduced in Hermes 2 Pro and continuously refined since. Q: Should I use Hermes Agent or OpenClaw? A: Choose Hermes Agent if you need automatic skill generation and broad messaging platform support. Choose OpenClaw if you want a mature MCP ecosystem and a proven production track record. Q: Can Hermes handle Japanese? A: Natural Japanese text generation is limited. However, for Function Calling and JSON output use cases, it is practical. For high-quality Japanese generation, Qwen 3.5 or Gemma 4 are recommended. Q: Is the 512K context of Hermes 4.3 actually useful? A: Yes — it enables whole-codebase comprehension, long-document Q&A, and extended conversation memory, all of which are valuable in real-world deployments. Q: Can I run Hermes without a GPU? A: Hermes 3 3B at Q4 quantization can run on CPU only, but response speed will be significantly reduced. A GPU is recommended for practical use. Q: Which model is recommended? A: With 24 GB VRAM, Hermes 4.3 36B (Q4) is the best choice. With 8–10 GB, Hermes 3 8B offers the best balance of performance and accessibility.

Oflight provides end-to-end support for deploying Hermes and other local LLMs in production — from Function Calling implementation and Hermes Agent integration to on-premises AI agent architecture. Whether you're evaluating models or building a full agentic system, our team covers model selection through deployment. Learn more about our AI consulting services