Nous Research

Nous Research is an AI lab known for the Hermes family of open-weight models, for Hermes Agent, an open source agent that runs on your server, and for the Psyche distributed training network.

The Hermes models are described in the Hermes 4 technical report as neutrally aligned generalist models that combine reasoning with instruction following. The weights are on Hugging Face, in formats such as BF16, FP8 and GGUF.

Hermes Agent, according to the official site, is an autonomous agent that runs on your server, remembers what it learns and uses the MIT license. It works with several model providers, including Nous Portal, a gateway from Nous Research itself that brings more than 300 models behind a single login.

Psyche is the company's distributed training network. Nous Research says Hermes 4.3 was its first production model post-trained entirely on it.

What it is for: teams that want open models they can run locally and an agent they host themselves, instead of depending only on a closed API.

Nous Research is for developers and technical teams that want to run open-weight models on their own, host an AI agent on their own server or follow open research on distributed training.

The usage profiles that the official sources support are these:

• Teams running models locally: the Hermes weights are on Hugging Face in BF16, FP8 and GGUF, and the model pages include instructions for running the GGUF versions with Hermes Agent.
• Teams that want a self-hosted autonomous agent: Hermes Agent installs with one command on Linux, macOS and WSL2, through PowerShell on Windows, and there is a desktop app for macOS and Windows.
• Teams using several model providers: Hermes Agent accepts API providers such as DeepSeek and OpenAI, besides Nous Portal, and switches provider with the hermes model command.
• Teams researching distributed training: the Psyche network and the Hermes 4 technical report are published.

Prerequisites: a server or computer to run the agent, and a model provider account for it to use. The macOS desktop app works only on Apple Silicon.

The main features of Nous Research are the Hermes family of open-weight models, Hermes Agent, Nous Portal and the Psyche distributed training network.

• Hermes family: neutrally aligned generalist models, according to the Hermes 4 technical report, which integrates self-reflective reasoning with broad instruction following. The weights are on Hugging Face in BF16, FP8 and GGUF.
• Hermes 4.3: according to Nous Research, it was post-trained entirely on the Psyche network, on top of ByteDance's Seed 36B base model, and matches Hermes 4 70B at half the parameters.
• Hermes Agent: an open source, self-hosted autonomous agent, under the MIT license. It orchestrates several agents through its profiles system and keeps memory of what it learns.
• Nous Portal: a gateway with OAuth login, which the documentation says gives access to more than 300 models.
• Psyche: a distributed training network that uses the DisTrO optimizer and records consensus state in a smart contract on Solana, while gradients travel over its own P2P mesh.

For the current version of each model, the Nous Research releases page is the reference.

Nous Research is worth it for teams that want open models they can download and run, and an agent they host themselves. The limits lie in setting up your own model provider and in depending on third-party base models.

Pros:

• Open weights of the Hermes models on Hugging Face, in BF16, FP8 and GGUF, which allows running locally.
• Hermes Agent uses the MIT license and runs on your own server.
• Hermes Agent works with several model providers, without tying you to Nous Portal.
• The Hermes 4 technical report is public, and the company itself describes Psyche in detail.
• Installation takes one command on Linux, macOS, WSL2 and Windows, plus a desktop app.

Cons:

• To use Hermes Agent you must set up a model provider, through Nous Portal or your own API keys.
• The Hermes Agent desktop app for macOS works only on Apple Silicon.
• Hermes 4.3 was built on a ByteDance base model, so it depends on a third-party model.
• Nous Portal brings together models from other labs, so what it offers depends on those suppliers.

The main alternatives to Nous Research are other labs that publish open-weight models.

• Meta, with the Llama family: open-weight models in several sizes, published by Meta itself.
• Mistral AI: a European lab that publishes part of its models with open weights.
• Alibaba, with the Qwen family: open models in several sizes and modalities.
• Liquid AI: small open models aimed at running on devices, with the LEAP platform for deployment on iOS, Android and desktop, according to the company's documentation.
• DeepSeek: a lab that publishes reasoning models with open weights.

What sets Nous Research apart is combining, in one lab, open models, a self-hosted agent and a distributed training network. Teams that only need an open model to download find options at every lab in the list. Teams that also want an agent they host themselves, under the MIT license, and want to follow distributed training research find both on the Nous Research site.