Nous Hermes 2: The Community Sweet Spot

What Is Nous Hermes 2?

Nous Research is a small, community-oriented AI lab that built its reputation by asking a deceptively simple question: can we fine-tune open-weight base models better than the base model labs themselves? For the Hermes series, the answer has consistently been yes.

Nous Hermes 2 is a family of instruction-tuned models fine-tuned on a large, carefully curated dataset that Nous Research assembled from synthetic and human sources. The training data — often referred to informally as the "Nous dataset" — emphasizes reasoning depth, instruction diversity, and what the team calls "thoughtful alignment": safety behaviors that reflect actual harm potential rather than blanket topic refusals.

The practical result is a model that feels more like a knowledgeable colleague than a corporate chatbot. It will discuss topics that trigger ChatGPT's refusal filters, provided the context is reasonable. It also scores notably well on MT-Bench's multi-turn coherence tests, where maintaining conversational context across a long exchange matters more than raw knowledge recall.

Model Variants: Which One to Use

Nous Hermes 2 is not a single model — it's a family of fine-tunes across different base models and sizes. Choose based on your available hardware and use case.

Understanding the Alignment Approach

One of the most common questions about Nous Hermes 2 is where it sits on the spectrum between "refuses everything" and "no guardrails at all." The answer matters for legitimate professional use cases.

Nous Hermes 2 lands in a deliberate middle zone. It will discuss security topics, legal gray areas, mature content in appropriate contexts, and competitive intelligence without the reflexive hedging ChatGPT applies. It will not assist with clearly harmful requests. The Nous team describes this as "not treating users like children while still having a conscience."

For most professional use cases — writing, research, analysis, coding, summarization — the alignment difference from ChatGPT is invisible. You only notice it when you're doing work that triggers ChatGPT's overly cautious filters: security research, competitive analysis, medical professional queries, fiction with mature themes, or frank discussion of controversial topics.

Benchmark Performance

Two numbers stand out: the Nous Hermes 2 Yi 34B hits 74.8% MMLU — higher than GPT-3.5 Turbo and within striking distance of Llama 3.1 8B, on a fully local, Apache 2.0 model. The 7B variant's 7.7 MT-Bench score is above Zephyr and Dolphin and competitive with OpenChat, despite using an older Mistral base. Nous's fine-tuning genuinely moves the needle.

Capability Profile (7B Variant)

Honest Strengths and Weaknesses

Deployment: Ollama Quickstart

The fastest path to running Nous Hermes 2 is via Ollama. All four major variants are available in the Ollama library. See our Quickstart guide for Ollama installation instructions.

# 7B — default, runs on 8GB RAM
ollama run nous-hermes2

# 13B — better quality, needs ~16GB RAM
ollama run nous-hermes2:13b

# 34B Yi — long context flagship, needs ~32GB RAM
ollama run nous-hermes2-yi-34b

# Mixtral MoE — needs ~32GB RAM, high quality
ollama run nous-hermes2-mixtral

ChatML Prompt Format

Nous Hermes 2 uses the ChatML format — the same format used by OpenAI's API. This is a significant practical advantage: any code written for the OpenAI API needs minimal changes to work with Nous Hermes 2 via an Ollama-backed OpenAI-compatible endpoint.

<|im_start|>system
You are a helpful AI assistant specialized in data privacy law.<|im_end|>
<|im_start|>user
What are the key differences between GDPR and CCPA?<|im_end|>
<|im_start|>assistant

import requests, json

OLLAMA_URL = "http://localhost:11434/api/chat"

def chat(messages: list, model: str = "nous-hermes2") -> str:
    payload = {
        "model": model,
        "messages": messages,
        "stream": False,
        "options": {
            "temperature": 0.7,
            "num_ctx": 8192,
        }
    }
    r = requests.post(OLLAMA_URL, json=payload)
    r.raise_for_status()
    return r.json()["message"]["content"]

# Example: privacy policy analysis
messages = [
    {
        "role": "system",
        "content": "You are an expert in data privacy compliance."
    },
    {
        "role": "user",
        "content": "Summarize the user rights under GDPR in plain English."
    }
]

response = chat(messages)
print(response)

Best Use Cases

Getting the Most Out of Nous Hermes 2

Nous Hermes 2 responds well to specific system prompts. Unlike ChatGPT, where system prompts are often overridden by safety layers, the system prompt in Nous Hermes 2 is genuinely authoritative — the model takes it seriously throughout the conversation.

Effective System Prompt Patterns

You are an expert [domain] professional assisting [role description].
Provide direct, detailed responses. Assume the user has professional
background in this field. Avoid excessive disclaimers.

You are a careful document analyst. When given documents, extract
key information accurately. Cite specific sections when making claims.
If information is absent from the document, say so explicitly rather
than inferring.

You are a structured data extraction assistant. Always respond with
valid JSON matching the schema provided by the user. Do not include
explanatory text outside the JSON object.

When to Choose Something Else