Nous Research

The Hermes family ships as downloadable weights, not just a hosted endpoint. That means you can run the models on your own hardware, fine-tune them, or inspect how they behave. For teams that need data to stay in-house, this is the difference between using a model and owning one. No permission needed.

Hermes Agent is a self-hosted framework that runs on your own server and tool set. It keeps persistent memory across sessions and can store reusable skills, so a workflow that took twenty tool calls the first time around often shrinks to a handful once the agent has seen the problem before. It works with many model providers, so you aren't tied to one vendor for the underlying model.

Nous Portal gives programmatic access to the models through an OpenAI-compatible interface. Already have code written against OpenAI's client? Switching is mostly a base URL and key change. This route skips the hardware setup that self-hosting demands.

The lab works on coordinated training across distributed hardware, including the Psyche Network and the DisTrO optimizer, which compress gradients and update asynchronously so that large training jobs can run across machines that aren't sitting in one data center. The practical payoff is lower communication overhead between nodes. Less bandwidth, more reach.

YaRN is a method for extending how much text a model can handle at once. It has been cited widely and adopted by other labs, which says something about how well the approach travels beyond its origin. For users, longer context means fewer awkward splits when you're feeding in long documents, code files, or transcripts that would otherwise need to be chopped into pieces.

Research notes, model cards, and training details are published alongside each release, which lets outside teams verify claims, reproduce results, and rebuild setups without guessing at what happened behind closed doors. It's a different posture from labs that only describe results in marketing terms.