Hermes 4 is the open-weight model family from Nous Research that adds hybrid reasoning to the Hermes agentic recipe: a single model can answer instantly or stop to think with <think> tags before replying. It was released in August 2025 in three sizes and keeps the tool calling that made the series famous. In this guide you will see what changed since Hermes 3, how thinking mode works, its capabilities for agents and how to run it on your own machine. The same explanation is available in Spanish.

Key takeaways

  • Hermes 4 is a family of open-weight models Nous Research released on 25 August 2025, in three sizes: 14B, 70B and 405B parameters.
  • Its headline change is hybrid reasoning: the model decides whether to answer directly or open a <think>…</think> block to deliberate, something Hermes 3 did not offer.
  • It remains an agentic model: it uses the Hermes tool format (<tool_call>) and ships automatic parsers built into vLLM and SGLang.
  • Each size starts from a different base: the 14B from Qwen3-14B (Apache-2.0 licence), and the 70B and 405B from Llama 3.1 (Llama3 licence), with 131K tokens of context on the 70B.
  • In reasoning mode, the 405B posts high scores: 96.3% on MATH-500, 70.5% on GPQA Diamond and 81.9% on AIME 2024, per Nous’s technical report.

What is new in Hermes 4?

Hermes 4 is the fourth generation of the fine-tuned model series from Nous Research, a collective that takes open base models and retrains them to follow instructions, use tools and answer without the restrictions of commercial versions. The difference from Hermes 3 is not a new pretraining run but the fine-tuning: Hermes 4 is pure post-training on top of existing base models.

The family comes in three sizes aimed at different hardware budgets. The 14B is the lightest and starts from Qwen3-14B under the Apache-2.0 licence, the most permissive of the three. The 70B and 405B start from Llama 3.1 and inherit its licence; the 405B adds around 406 billion parameters and is the flagship. The 70B offers a 131,000-token context window, enough for long documents and extended conversations.

Nous also greatly expanded the fine-tuning data. The post-training corpus went from Hermes 3’s roughly one million samples and 1.2 billion tokens to about 5 million samples and 60 billion tokens, blending data with and without reasoning. That jump in scale is what underpins the model’s main new feature.

How does hybrid reasoning and thinking mode work?

Hybrid reasoning means a single model does two things that used to require separate models. When the question is simple, it answers directly, without spending extra tokens. When the task is complex (maths, code, logic), it opens a deliberation block delimited by <think> and </think> tags, reasons step by step inside it and only then writes the final answer. It is the same idea popularised by models like DeepSeek-R1, but built into an open model with a switch.

The behaviour is controlled from code. In the chat template, a thinking=True flag (or an instruction in the system message) activates deliberation; with keep_cots=True you decide whether to keep or discard the chain of thought in the output. So you can use the same model in fast mode for a chatbot and in reflective mode for an agent solving hard problems, without changing weights or server.

That reflective mode is what drives the 405B’s numbers: 96.3% on MATH-500, 78.1% on AIME 2025 and 61.3% on LiveCodeBench, all measured with reasoning on. Nous also reports a striking figure on RefusalBench, a test of undue refusals: the 405B scores 57.1%, well above GPT-4o’s 17.67% and Claude Sonnet 4’s 17%. According to Nous Research, Hermes 4 aims to be a "neutrally aligned" model that obeys the user rather than rejecting legitimate requests.

What agentic and tool capabilities does it offer?

The trademark of the Hermes series is tool calling, and Hermes 4 keeps it intact. The format is XML: the signatures of available functions are described inside <tools> tags in the system message, and when the model decides to use one, it emits a <tool_call> object with the name and arguments. This scheme is identical to Hermes 3, so agent code written for the previous version works unchanged.

A snippet of a tool call looks this simple:

<tool_call>
{"name": "get_weather", "arguments": {"city": "Madrid"}}
</tool_call>

The interesting part is how it combines with hybrid reasoning: the model can think inside a <think> block, decide which tool it needs and emit the call, all in the same turn. So you do not have to write the parser by hand, Hermes 4 ships tool parsers built into vLLM and SGLang, the two most common inference engines for open models. That integration is what makes Hermes 4 a comfortable base for agents that query APIs, run code or read databases. If you come from function calling with Ollama, you will recognise the pattern, though the Hermes XML format is its own.

How do you run Hermes 4 on your own machine?

Being open-weight, you can download Hermes 4 from Hugging Face and serve it yourself. The recommended route for production is vLLM, because it includes the Hermes tool parser out of the box. Serving the 70B variant with tool calling enabled is a single command:

pip install vllm

# Serve Hermes 4 70B with the Hermes tool parser
vllm serve NousResearch/Hermes-4-70B \
  --tool-call-parser hermes \
  --enable-auto-tool-choice

Size drives the choice. The 405B needs several data-centre GPUs even in the FP8 quantised version Nous publishes, so for most people the 70B or 14B are the realistic option. For anyone who wants to try it with little effort, the LM Studio community keeps quantised GGUF versions that run on a powerful laptop. If you want maximum serving speed in production, it is worth reading how to serve an LLM with vLLM; if you prefer the simplest path on your own machine, start by installing Ollama and a GGUF variant. On OpenRouter, the 70B runs around $0.13 per million input tokens and $0.40 for output, a handy cost reference if you would rather not host it.

How does Hermes 4 differ from Hermes 3?

The comparison with Hermes 3 makes the evolution clear. Hermes 3 was a good instruction and agentic model, but it always answered in one go, with no explicit reasoning phase. Hermes 4 adds exactly that: thinking mode with <think> tags, which lifts performance on maths, code and logic tasks without giving up fast answers when they are not needed.

There are other differences worth keeping in mind. Hermes 3 was built only on Llama; Hermes 4 adds a variant on Qwen3 (the 14B), which broadens the licensing options. The volume of fine-tuning data grew fivefold, and the scores on reasoning tests rise noticeably. What does not change is the tool format or the neutral-alignment philosophy, so moving from Hermes 3 to Hermes 4 is more an upgrade than a rewrite. If you have never used the series, Hermes 4 is today the recommended entry point.

Frequently asked questions

Is Nous Hermes 4 free and open source?

Hermes 4’s weights are open and free to download from Hugging Face, but the licence depends on the size. The 14B, based on Qwen3-14B, uses the permissive Apache-2.0 licence. The 70B and 405B start from Llama 3.1 and inherit the Llama3 licence, which imposes some conditions on large-scale use. Check the licence of the specific size before a commercial deployment.

What hardware do I need to run Hermes 4?

It depends on the size. The 14B fits on a single consumer GPU with quantisation and is the most accessible option. The 70B calls for a high-end GPU or several, and the 405B needs several data-centre GPUs even in the FP8 version. To try it without buying hardware, the simplest route is a GGUF variant with Ollama or paying per token at a provider like OpenRouter.

How does hybrid reasoning differ from a normal reasoning model?

A pure reasoning model always deliberates before answering, which spends more tokens and time even when the question is trivial. Hermes 4’s hybrid reasoning lets you choose: you turn on thinking mode with a flag when the task deserves it and turn it off when you want instant answers. A single model covers both cases, without keeping separate weights.

Conclusion

Hermes 4 is the update of one of the most AI-agent oriented open model series. It adds hybrid reasoning with <think> tags, keeps the Hermes tool calling and comes in three sizes (14B, 70B and 405B) with parsers ready for vLLM and SGLang. Its reasoning-mode numbers (96.3% on MATH-500, 70.5% on GPQA) put it on par with far more closed models. The next step is to download it from Hugging Face, serve the 70B with vllm serve and give it a tool to see the agentic loop in action.

Sources: [1] Official Hermes 4 site (Nous Research)[1], [2] Hermes-4-405B model card on Hugging Face[2], [3] Hermes 4 70B on OpenRouter (pricing and spec)[3], [4] Release coverage at MarkTechPost[4].

Visitas: 0

Sources

  1. Official Hermes 4 site (Nous Research)
  2. Hermes-4-405B model card on Hugging Face
  3. Hermes 4 70B on OpenRouter (pricing and spec)
  4. Release coverage at MarkTechPost

Route: Self-hosted Agentic Models and Tool Calling