Nous Research has introduced Hermes 3, a new family of language models designed around a clear idea: the system prompt should strongly shape how the model behaves. According to the technical report, the models are built for high controllability and neutral alignment, with the goal of following instructions closely and adapting to the world view specified by the user or developer.
The release includes Instruct models with 8, 70, and 405 billion parameters. All are based on Meta's open-source model Llama 3.1, and Nous Research says Hermes 3 performs strongly among models with open weights.
What Hermes 3 Is Designed To Do
The central claim around Hermes 3 is not simply that it can answer questions or generate text. Nous Research presents the model family as a controllable system that can be steered through instructions, especially through the system prompt.
That matters because the system prompt is where developers usually define a model's role, tone, limits, and operating assumptions. In the source material, Hermes 3 is described as adapting to the world view specified there, rather than applying a separate moral filter in the way some proprietary commercial models may do.
The report frames this as a contrast with models that may refuse instructions for moral reasons. For Hermes 3, the report says there is no "latent thoughtcrime." The point, as presented, is that the model should not silently reject a requested perspective merely because it differs from a built-in preference.
This does not make Hermes 3 a single-purpose model. The source describes it as a broad family of Instruct models intended to handle several kinds of language tasks, from reasoning to structured output.
Model Sizes, Base Model, And Training
Hermes 3 comes in three Instruct sizes: 8, 70, and 405 billion parameters. The parameter range gives the family multiple deployment profiles while keeping the same overall design direction.
The models are based on Meta's open-source model Llama 3.1. That base is important to the article's subject because Nous Research compares Hermes 3 not only with other open-weight models, but also with the underlying models from Meta.
The training process described in the source has two phases. First came supervised fine-tuning, or SFT. Then came a phase using Direct Preference Optimization, or DPO.
Nearly 400 million tokens were used for the SFT phase. The models were evaluated epoch-wise, and the best checkpoints for the 8B and 405B models were selected. In plain terms, the training process included repeated evaluation points, with selected versions kept when they performed best under the chosen criteria.
The source also says the training mix included synthetically created reasoning tasks and expressive applications. Those expressive applications included role-playing and creative writing, which helps explain why the model family is presented as both task-oriented and flexible in style.
Benchmarks And Capabilities
Nous Research says Hermes 3 achieves top scores in several public benchmarks among models with open weights. The named benchmarks are ARC, BoolQ, HellaSwag, IFEval, and Winogrande.
The source also says Hermes 3 outperforms Meta's Llama 3.1. That comparison is especially relevant because Hermes 3 is based on Meta's open-source model Llama 3.1, so the claim is that the additional training and alignment approach improved results over the underlying models.
The listed skills go beyond basic chat completion. According to Nous Research, Hermes 3 handles:
- reasoning
- reward modeling
- "scratchpads" for intermediate results
- structured output with XML tags
- generation of internal monologues for transparent decision-making
- Mermaid diagrams for visual communication
Those capabilities point to a model family intended for workflows where format matters. Structured output with XML tags can help when responses need to be parsed or organized. Mermaid diagrams can help turn information into visual communication. Scratchpads and internal monologues, as described by Nous Research, relate to exposing intermediate reasoning or decision steps.
The models can also use external tools. In addition, the source says they can cite information from documents through Retrieval Augmented Generation to answer questions. That places Hermes 3 in the category of models meant to work with external context, rather than only relying on what is already contained in the model itself.
Why Neutral Alignment Is The Core Claim
The strongest theme in the Hermes 3 release is not a single benchmark result. It is the combination of controllability and neutral alignment.
For developers, controllability means the model should be easier to direct. If a prompt asks for a specific format, viewpoint, or task behavior, the model is intended to follow that instruction precisely. The source frames this as a defining feature of Hermes 3.
Neutral alignment is presented as the companion idea. Rather than embedding a fixed moral stance that may override the user's instructions, the model is described as adapting to the world view specified in the system prompt.
That positioning sets Hermes 3 apart from proprietary commercial models in the source article. The difference is not described as size alone, or benchmark performance alone, but as an approach to how the model should respond when instructions define its operating frame.
Availability And The Open-Weight Context
The Hermes 3 models are available on Hugging Face. That availability fits with the source's repeated focus on open weights and public benchmark comparisons.
For the open-weight model community, the release is notable because it combines several elements in one package: multiple model sizes, a base in Meta's open-source model Llama 3.1, a two-phase training process, strong benchmark claims, tool use, Retrieval Augmented Generation, and a clear emphasis on system-prompt controllability.
The practical takeaway is straightforward. Hermes 3 is being positioned as a language model family for users who want strong instruction following, flexible behavior, and structured outputs, while keeping the model's alignment closer to the prompt-defined role than to a fixed refusal pattern.