Hermes

Key Features

• Neutral Alignment: This is the idea at the heart of Hermes. Nous describes its models as “neutrally-aligned and steerable.” In plain terms: the model follows your instructions and system prompt, not a company’s ethics code. Nous built its own test to measure this, called RefusalBench. It uses 166 prompts across 32 categories to see how often a model refuses to answer. Hermes 4 scored 57.1% in reasoning mode. GPT-4o scored about 17.7%, and Claude about 17%. A higher score means fewer refusals. Whether that is good or bad depends entirely on what you are building.
• Reasoning You Can Switch On and Off: Hermes 4 can think before it answers, or answer straight away. You choose. When thinking is on, the model works through the problem inside <think>...</think> tags. When it is off, you get a faster, cheaper reply. This is useful in agent workflows. You can run quick, cheap steps for routine tasks, then switch thinking on only for the hard step.
• An Agent That Learns: Hermes Agent has a built-in learning loop. It creates skills from experience, improves them while it works, reminds itself to save what it learns, and builds a picture of you across sessions. The problem it solves is agent amnesia. A normal agent finishes a task, then forgets everything. Hermes turns the task into a reusable skill, saves it, and recalls it next time. The skill library grows instead of resetting.
• The Training Flywheel: Here is the part most reviews miss. Hermes Agent records what it does and feeds that data into Atropos, the reinforcement learning framework built by Nous. That data can then train the next generation of models. So a loop forms: the product creates data, the data creates a better model, and the better model creates a better product. Very few vendors have built this loop so openly. It is worth understanding before you deploy.
• Runs Anywhere, Talks Everywhere: Hermes Agent is not stuck in your code editor or on your laptop. It runs on six different backends, including Docker, SSH and serverless options like Modal. Serverless instances sleep when idle, so an always-on agent costs almost nothing between tasks. On the front end, one gateway connects to more than 20 platforms: Telegram, Discord, Slack, WhatsApp, Signal, Teams, Email, SMS and others. You can message it from your phone while it works on a server you never log into.
• Skills, Tools and MCP: The agent comes with more than 60 built-in tools. It connects to any MCP server for more. Skills follow the open agentskills.io standard, so they are portable and shareable, and the community contributes them through a Skills Hub. A skill you write is not locked to Nous. The agent can also split work across sub-agents and run them in parallel.
• Open Weights, Open Everything: Hermes 4 was released as open-weight models in three sizes: 14B, 70B and 405B. They are built on Meta’s Llama 3.1. Nous published the full weights and the training methods. Hermes Agent itself is MIT-licensed on GitHub, which means you can modify it, build a commercial product on it, and keep your changes private with no licence fee.

The more important point for enterprises is what these releases make possible. Because you can download and run everything yourself, your data never has to leave your infrastructure. If your organisation must keep data in-country, or must run offline, Hermes becomes a more competitive option to be considered.

Company Background

Nous Research started in 2022. A group of AI researchers and engineers began working together informally on Discord, GitHub and Twitter. That origin still shows in the products today. The company is now based in New York. Its core team includes CEO Jeffrey Quesnelle, co-founder Karan Malhotra, and a pseudonymous researcher known as “Teknium,” who created the original Hermes line and leads post-training.

The funding story is unusual. Nous first raised around $20 million in seed rounds. Then, in April 2025, it raised a $50 million Series A. The round was funded almost entirely by Paradigm, a crypto-focused venture firm. The valuation was reported at $1 billion, but as a token valuation, not a normal equity valuation. Total funding is around $65 million. Paradigm’s reason for investing was the contrast itself: an open, community-driven lab set against closed, centralised ones.

The crypto side is real, and you should understand it. Alongside its models, Nous is building Psyche. This is a decentralised training network. It links together GPUs owned by many different people, coordinated through the Solana blockchain, so a model can be trained across the internet rather than inside one data centre. Co-founder Malhotra has been open about the scepticism this attracts. He says Nous chose to prove its AI research first, and add the blockchain second, not the other way around. You may or may not like the token model. But the research has been strong enough that the question can be set aside when judging the products.

User Experience

• Hermes Agent: Installation really does take about a minute. There is a desktop installer for Windows and macOS, and a single command line for Linux, macOS and even Android. One command, hermes setup --portal, gets you a working agent with one login. That login covers the model plus web search, image generation, text-to-speech and a cloud browser. The feel is different from other tools. It is less “a chatbot in a window” and more “a process you deploy somewhere, then message.” If you are coming from Copilot or Cursor, this takes some getting used to.
• Hermes Cloud: If you do not want to manage servers, Nous will host the agent for you. Hermes Cloud runs your agent on a dedicated instance with its own dashboard. You deploy in seconds and pay by the hour. When you stop the instance, it drops to storage-only pricing. An idle agent costs a few cents a day.
• Nous Chat and the “Orb”: Nous built its own chat interface, and it deliberately avoids making the AI feel like a person. Your memory is stored in an “orb.” You fill the orb with what you want the AI to remember, then carry it across different models and prompt templates. Different workspaces get different orbs. Nous describes the goal as talking to a directory of experts rather than to one fixed character. It is a deliberate move against the industry trend of a single assistant personality.
• For Developers: Hermes works with Nous Portal, OpenRouter, OpenAI, or any compatible endpoint. You are not locked in to Nous for inference. Nous Portal bundles 300+ models under one subscription, so you do not need five separate API keys for the model, search, images, speech and browsing. The documentation is thorough. It even publishes machine-readable files (llms.txt) so other AI coding agents can read it directly.
• Known Friction Points: The neutral-alignment stance is the big question, and it deserves an honest answer rather than praise or panic. A model tuned to refuse less is the right tool for red-team testing, security research, and any application where a refusal is a product failure. It is the wrong tool for a customer-facing enterprise system, where one bad output becomes a headline. There is also a benchmark caveat. Artificial Analysis ranked the 405B model fairly low, but only because they tested it in non-reasoning mode. On raw capability, the community view is that larger open models such as DeepSeek R1 are still ahead. The honest strategic read is this: Hermes makes sense when two things are both true. You need the steerability that standard alignment removes, and you need to run open weights on your own infrastructure. If neither is true, the extra work of running open weights will not pay off. Finally, there is a governance question, as we discussed in When AI runs its own Retrospective. A self-improving agent with persistent memory and a changing skill library is a moving target. The agent your security team reviewed on Monday is not the same agent running on Friday.

Cost

Nous prices in line with its open-source philosophy. The software is free. The weights are free. You pay only for the compute and tools you actually use. There is no per-seat subscription blocking the door.

Free Tier

• Hermes Agent (Open Source) – MIT-licensed, available on GitHub.
  • Price: $0 in licence fees.
  • Includes: the full agent, the learning loop, skills, memory, 60+ tools, MCP support, and the messaging gateway.
  • Note: you still pay for inference from whichever model provider you choose.
• Open-Weight Models – Hermes 4 (14B, 70B, 405B) on Hugging Face.
  • Price: $0 in licence fees; you pay for compute.
  • Use case: data-sovereign deployments, offline environments, and regulated industries that need on-premises inference. The smaller models run on modest hardware. The 405B does not.

Hermes Cloud (Hosted Agent, Billed Hourly)

Prices come out of your account credit each hour. They exclude inference and tool usage. You need at least $2 of credit to start.

• Small – Light personal use. 5 sessions, 1GB RAM, 2 vCPUs.
  • Running: ~$0.32/day · Stopped: ~$0.06/day
• Medium – Regular work. 10 sessions, 2GB RAM, 4 vCPUs.
  • Running: ~$0.59/day · Stopped: ~$0.06/day
• Large – Bigger projects. 20 sessions, 4GB RAM, 8 vCPUs.
  • Running: ~$1.12/day · Stopped: ~$0.06/day

“Running” means compute plus storage while the instance is on. “Stopped” means storage only, which keeps your data between sessions. So an agent you use now and then costs about the price of one coffee per month to keep alive.

Inference and Tools (Pay As You Go)

• Hermes 4 405B (via OpenRouter): about $1 per million input tokens and $3 per million output tokens, with a 131K context window.
• Nous Portal: one subscription covers 300+ models plus the tools, i.e., web search, image generation, text-to-speech and a cloud browser. No extra accounts needed.
• Example tool costs: search credits from about $0.0005 each; cloud browser about $0.0011 per minute; image generation from about $0.005 to $0.26 per image, depending on the model.

For current prices, see the Nous Portal pricing page.

---

In summary, Hermes is not trying to be a friendlier ChatGPT. If you judge it that way, you will reach the wrong conclusion. It is a deliberate alternative. Open weights instead of closed APIs. Alignment to the user instead of to the company. Distributed training instead of giant data centres. And an agent that builds up knowledge instead of forgetting everything each time.

That makes it a strong choice for a specific group: researchers, security teams, developers with data-sovereignty rules, and anyone frustrated by a model that refuses reasonable requests. It makes it a poor choice for a regulated, customer-facing service, where a low refusal rate becomes a risk rather than a benefit.

But there is one reason to watch Hermes no matter which group you are in. It is the self-improving loop. An agent that writes its own skills, rewrites them as it works, and feeds its own experience back into training the next model is not static software. You cannot review it once and sign it off. That is the future arriving early. And it is exactly the case that will test whether your governance model is ready.

Related reading: When AI Runs Its Own Retrospective – on recursive self-improvement, and what it does to the human checkpoint.