Hermes and Nous Research: who makes it and how it works
Last updated: 27 October 2026
"Hermes" is two things from the same team: a family of open AI models, and an AI agent program called Hermes Agent. Both come from Nous Research. This page explains who they are and how the pieces fit, using only Nous Research's own pages.
Who is Nous Research?
Nous Research describes itself as a lab for open AI training and research. It says it trains models, builds agents and builds the tools needed to run them. Its stated mission is "to create and proliferate open access to intelligence." In plain words: it wants strong AI that anyone can download and run, not only big companies.
Its site lists its main research areas as agents, model design, making training data, fine-tuning (extra training on top of an existing model) and reasoning. Its models are published on Hugging Face, a public site where AI models are shared, and its code is on GitHub.
Hermes, the models
A model is the AI "brain" that reads text and writes text. Nous Research's Hermes models are open-weight. The weights are the numbers the model learned during training, and anyone can download them and run the model on their own hardware.
Nous does not usually train these models from zero. It takes a base model from another lab and trains it further. For example, its model card says Hermes 4.3 36B is built on ByteDance's Seed 36B base model. "36B" means about 36 billion learned numbers. The Hermes 4 family also comes in 14B, 70B and 405B sizes.
Per the Hermes 4.3 model card, the models can:
- Think before answering, if you ask. In "reasoning mode" the model writes its working inside
<think>tags, then gives the answer. You can switch this off for faster replies. - Call tools. You describe a tool, such as "get the weather for a city," and the model replies with a structured request to use it. This is what lets a model act as part of an agent.
- Return tidy data. It is trained to produce valid JSON, a common format programs use to swap data.
Hermes Agent, the program
Hermes Agent is the program wrapped around a model. It gives the model memory, tools and places to talk to you. It is open-source, and it can work with Nous's own models or with other providers such as OpenRouter or OpenAI.
Its docs call it "self-improving." In plain words, three things happen as you use it:
- When it solves a task, it can save the steps as a reusable skill and improve that skill later.
- It keeps memory across conversations, including notes about you.
- It can search its own past chats to recall what was said before.
The docs say it can run on a cheap rented server or a cloud service, not just your laptop. You can reach it from the terminal or from chat apps such as Telegram, Discord or Slack. It can run scheduled jobs, start helper agents to work in parallel, and connect to outside tools through MCP (Model Context Protocol, a common way to plug tools into AI programs).
Where does blockchain fit in?
Hermes Agent itself does not need a blockchain. The link is in how Nous trains some of its models.
Training a big model normally needs thousands of chips in one building. Nous built Psyche, a network that spreads training across computers in different places over the internet. Its own write-up says the coordinator, the part that keeps track of who is taking part and what step the training is on, "lives on-chain in a smart contract" on the Solana blockchain. A smart contract is a program stored on a blockchain that anyone can check. Nous says this makes the network harder to shut down or censor.
The Hermes 4.3 model card says it is the first Hermes model trained this way, over the internet using Psyche.
This is not NEAR, and nothing here involves FLTC. It is a separate example of a blockchain used to coordinate work instead of moving money.
What people use it for
- Running an open model on your own hardware, so your chats are not sent to an outside company.
- A personal assistant that remembers your projects and can run tasks on a schedule.
- Building other software that needs a model able to call tools.
Honest limits
- Facts can be wrong. On SimpleQA, a short factual-questions test reported in the Hermes 4.3 model card, the 36B model scored 6.0. Do not trust it for facts, prices or addresses without checking.
- Fewer refusals cuts both ways. Nous says its models are built to refuse less and follow your instructions more. That means fewer built-in brakes. You are the one who has to say no to a risky action.
- Big models need big hardware. Running the larger sizes yourself needs strong graphics cards. Smaller, compressed versions exist but give up some quality.
- An agent can act, so it can make mistakes that matter. Hermes Agent has command approval and container isolation settings. Keep them on, and never give any agent your wallet's recovery phrase. See NEAR scams for why.
- Psyche checking is still unfinished. Nous says that checking every computer in the network did honest work is "still an open problem" in practice.
Sources
- Nous Research website: https://nousresearch.com/
- Nous Research on Hugging Face: https://huggingface.co/NousResearch
- Hermes 4.3 36B model card: https://huggingface.co/NousResearch/Hermes-4.3-36B
- Psyche network architecture (Nous Research): https://nousresearch.com/nous-psyche/
- Hermes Agent documentation: https://hermes-agent.nousresearch.com/docs/
- Hermes Agent on GitHub: https://github.com/NousResearch/hermes-agent