Nous Research and Hermes: Neutral Alignment Explained

Nous Research is an American open-source AI research lab that grew out of a Discord community around 2023. It makes the Hermes and OpenHermes models, one of the most widely used lineages among uncensored models. Its design rests on a philosophy the developers call "neutral alignment": rather than stacking heavy refusal behavior onto the model, Nous leaves the decision about how it behaves to the user's system prompt.

The lab also publishes its datasets, training recipes and a technical report for Hermes 3, so its releases can be checked against a public record. Its models have drawn tens of millions of downloads on Hugging Face and are known for following system prompts closely and returning structured JSON.

---

Who is Nous Research?#

Nous Research is an American AI research lab that took shape around 2023 from an open-source AI community on Discord. Its co-founders are Jeffrey Quesnelle, Karan Malhotra, Shivani Mitra and Teknium. It is better described as a leading research group in open-weight culture than as a rival to OpenAI or Anthropic.

The lab started as a community aiming to make the most of open-source models. Teknium, who works under an online handle, leads post-training (the additional training stage after pre-training) and has driven the OpenHermes line. The team's first output was data rather than a model: GPTeacher, released around March 2023, is a set of instruction-and-answer pairs generated by GPT-4. Good models needed high-quality training material at the time, and they released it for anyone to use. This data-first start is the root of the OpenHermes line.

Cumulative downloads of its models on Hugging Face reached the tens of millions by 2025, with estimates ranging from 33 million to over 55 million. In spring 2025 it also raised about $50 million in Series A funding led by Paradigm, at a reported valuation of roughly $1 billion. The lab supports open-weight culture at the research level: it publishes model weights (the learned numerical data) and distributes the recipe in a reproducible form.

What is neutral alignment in Hermes models?#

Neutral alignment, which the developers write as "neutrally-aligned," is the Hermes design philosophy of not stacking the content filters that large companies' models ship with. The decision about how the model behaves is left to the system prompt that the user writes. This policy applies to every Hermes generation.

A system prompt is the role and rules handed to the model before a conversation starts. A single line such as "act as X" shapes the model's persona, tone and judgment criteria. The technical report for Hermes 3, published in August 2024, states the philosophy clearly: Hermes 3 is designed as a general-purpose instruction and tool-use model that follows the system prompt faithfully and does not force an assistant persona when none is specified.

In the largest 405B version, an empty system prompt does not even produce a helpful general-assistant reply. The model is built to be that sensitive to the system prompt, so it does not impose a personality of its own.

What does this mean technically? In large commercial models, the behavior of declining is built up thickly during post-training, and refusing in borderline cases is a product of tuning toward human preferences, such as RLHF. Nous takes a different view: whether to decline or answer is just one of the countless behaviors learned in training, so its weight can be designed thin at the training-data stage.

Behind this is a critique by Teknium, who led OpenHermes development. He is reported to have described the excessive refusals of mainstream models as the world's "mode collapse." The term originally means a generative model losing output diversity and producing similar results. Here it expresses the concern that a model that refuses everything narrows the range of human inquiry.

This design matters most for roleplay, creative writing and persona building. Hermes puts instruction-following accuracy first, and steerability, meaning how faithfully the model follows a long custom system prompt and keeps a specified persona consistent, is the metric it emphasizes. Hermes treats the model as a tool, and "neutrally-aligned" also affirms low restriction as a legitimate option for how you use it.

How did the Hermes series evolve from OpenHermes to Hermes 4?#

Hermes grew from community-built training data into hybrid-reasoning models between 2023 and 2025. GPTeacher came first, followed by OpenHermes 2.5 with its one-million-sample dataset, Hermes 2 Pro with function calling, Hermes 3 on Llama 3.1, and Hermes 4 with hybrid reasoning around August 2025. The history is one of data as much as models.

In late 2023 two results appeared. Nous-Capybara used a data synthesis method called Amplify-Instruct to show competitive performance from only about 20,000 examples, and the original Hermes was a fine-tuned Llama. In parallel ran Teknium's OpenHermes line: OpenHermes 1 followed the original Hermes, and OpenHermes 2 chose the newly prominent Mistral 7B as its base. OpenHermes 2.5, released at the end of 2023, was the turning point. It was trained on about one million synthetic instruction samples (question-and-answer sets produced by a large model), and that dataset became the basis of the later Hermes 2 line.

In 2024 the lineage was folded into the Hermes name. Hermes 2, early in the year, produced a popular branch built on Mixtral 8x7B and tuned with DPO, a method that directly optimizes outputs toward human preferences. Hermes 2 Pro (Llama 3 8B base), in spring 2024, introduced function calling and structured output. Function calling is a mechanism by which the model calls an external tool or API in a fixed format. This moved Hermes from a model that converses well to one developers could use in real work, and its reputation for stable JSON output dates from this period.

Hermes 3, in August 2024, came in three Llama 3.1 sizes: 8B, 70B and 405B. It is unusual for a community model to publish a technical report in academic style. It is the milestone where neutral alignment was put in writing, and the recipe (data composition and training procedure) was shown with code in a reproducible form.

Hermes 4, around August 2025, made hybrid reasoning its headline feature. It lets the model switch between quick answers and longer, deliberate thinking depending on the task. Its low refusal rate in the large 70B and 405B versions drew attention again.

DeepHermes is a derivative that combines the Hermes recipe with reasoning training (learning to output the trace of thought). It was reported in early 2026 in 3B, 8B and 24B sizes.

How does Hermes differ from Dolphin, and which uses does it suit?#

Hermes and Dolphin both belong to the fine-tuning approach, but they differ in who makes them. Dolphin is associated with one person, Eric Hartford, while Nous Research publishes its datasets with code and a technical report for Hermes 3. Reproducibility is designed into Hermes from the start, which affects how easy it is to verify and how you use it.

Dolphin is a lineage associated with Eric Hartford as an individual, so the choices about what to mix in and what to leave out rest with a single craftsman. That gives the models a sense of unity, and the feeling of using someone's own work was part of the appeal. Nous Research, by contrast, treats Hermes as a product of research and development. Dolphin focused on the side that does the training, while Nous began with the training material itself.

huihui-ai works in the opposite order. It edits the weights of a finished standard model afterward, whereas Hermes does not stack the declining behavior thickly in the first place.

Hermes is repeatedly praised for its obedience to system prompts and for structured output, the ability to return a fixed format such as JSON reliably. A development style that publishes data and procedures and improves them together suits building features that fit specific uses deeply. The uses line up like this:

• Building a persona: Hermes follows long custom system prompts, so it suits roleplay and creative writing partners.
• Connecting to tools: function calling and structured output make it suitable as a component of an app.
• Sustaining long conversations: Hermes stays consistent, so it suits long-term work such as research notes.

The basis for choosing a builder is a record of what they did. Nous meets this standard at a high level: it published the OpenHermes 2.5 dataset of about one million samples and the Hermes 3 technical report. Choosing neutrality means taking responsibility for judgment yourself rather than leaving it to the maker. Picking a builder with data and technical reports on record helps you make that judgment.

What is Nous Research building today?#

As of 2026, Nous Research is shifting its focus from building models to building infrastructure. Its work now covers the DisTrO distributed-training optimizer, the Psyche distributed training network, the Atropos reinforcement-learning environment and the Forge agent framework. Hermes models remain published as open weights, so you can still run them yourself.

DisTrO is a mechanism that cuts the communication needed during training by orders of magnitude. The other three are a distributed training network, a reinforcement learning environment and an agent foundation, which extend the lab's research well beyond models themselves.

What is Deepent Research?#

Deepent Research is a browser-accessible LLM chat service built to refuse as little as possible. It keeps going with research that a standard AI would cut off mid-answer.

• It refuses nothing. We run uncensored and abliterated models with guardrails reduced as far as possible, so prompts a standard AI answers with "I can't help with that" still get a response (Qwen3.5 / Qwen3.8-27B-abliterated, depending on your plan).
• Your input is never used for training. What you look up is used only to provide the service, never to train AI or for any other purpose. Personal data is protected, and our privacy policy is public.
• Try it free. No GPU and no install, just open a browser. In the logged-out trial your conversation isn't saved, so you can start casually.

→ Try it free, no sign-up: https://deepentresearch.com

No question can stop research.

Frequently asked questions#

Is Nous Research an individual developer or an organization?#

⌄

Nous Research is a research organization, not an individual developer. It grew out of an open-source AI community on Discord around 2023 and is based in the United States. Its co-founders are Jeffrey Quesnelle, Karan Malhotra, Shivani Mitra and Teknium, who leads post-training. In spring 2025 it raised about $50 million in Series A funding led by Paradigm.

What is neutral alignment in the context of Hermes LLMs?#

⌄

Neutral alignment is the Hermes design philosophy of not stacking heavy content filters on the model and leaving its behavior to the user's system prompt. The developers describe the models as "neutrally-aligned." Instead of imposing a default assistant personality, Hermes 3 follows the system prompt closely, which makes it highly steerable for roleplay, creative writing and tool use.

What is the difference between OpenHermes and Hermes?#

⌄

OpenHermes is a line released under Teknium's personal name, and its dataset became the foundation of the organization's work. The OpenHermes 2.5 dataset holds about one million synthetic instruction samples. Since Hermes 2 in 2024, the lineage has been unified under the Hermes name as the organization's work.

What use cases are Hermes models best suited for?#

⌄

Hermes models suit roleplay, creative writing and app integration. They follow long custom system prompts closely and keep a specified persona consistent, which helps in character work and long conversations. They also handle function calling and structured output well, so they work as components inside applications that call external tools.

Disclaimer#

This article was drafted with AI assistance and then reviewed, fact-checked, and edited by the Deepent Research team before publication.
It is provided for educational and research purposes and is not intended to encourage unlawful activity or crime.
Where figures are cited, they come from third-party public reports and vendor-published values and may vary with measurement conditions.