Hermes Grows Up and Gets a Compliance Department

Remember when Nous Research was the group fine-tuning models in public, posting weights to Hugging Face, and arguing about distributed training on a timeline where the prevailing wisdom said frontier work required a hyperscaler’s cluster? That was the pitch: open models, open discourse, open weights. The Hermes line became shorthand for a certain kind of practitioner’s model, the one you ran yourself because you wanted to see what was under the hood.

That same group just confirmed a $1.5 billion valuation and shipped a product called “Hermes for Businesses.” The $90 million Series B was led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and 1789 Capital participating. Total funding now sits at $158 million. The stated use of proceeds is enterprise expansion, specifically letting companies deploy customized agents for multi-step workflows while keeping data private and secure.

I want to look at that last clause carefully, because it is doing more architectural work than it appears to.

Multi-step workflows are where agent architecture stops being theoretical

A single-turn model call is a function. You pass text in, you get text out, and if the output is wrong you can inspect the whole interaction in one screen. A multi-step workflow is something else entirely: a sequence of model calls where each step conditions on the results of the previous one, often mediated by tool use, retrieval, or external state mutation.

The failure modes compound differently. In single-turn inference, errors are independent. In a multi-step chain, errors are correlated and often amplifying, because step four is reasoning over step three’s hallucination as if it were ground truth. Anyone who has instrumented an agent loop in production knows the shape of this: the trace looks plausible at every individual hop, and the final answer is nonsense.

So when a company says it is selling customized agents for multi-step workflows, the interesting engineering questions are not about model quality in the benchmark sense. They are about:

• How state is represented and carried between steps, and whether that representation is inspectable
• Where the control flow lives — in the model’s own planning, in an external orchestrator, or split between the two
• What happens on step failure, and whether the system can roll back a side effect it already committed
• How customization is implemented, since fine-tuning, adapters, and prompt-level configuration carry very different maintenance costs
• Whether the trace is observable enough to debug after the fact

None of that is specified in the announcement, which is normal for a launch. But those are the axes on which an enterprise agent product actually succeeds or fails, and they are a different set of axes than the ones open-weight model releases are judged on.

Why the privacy framing matters more than it sounds

“Maintaining data privacy and security” is the kind of phrase that reads as boilerplate until you consider what it implies about deployment topology. If customer data cannot leave a customer’s boundary, the agent has to run somewhere that boundary permits. That pushes toward on-premise or customer-VPC deployment, which in turn pushes toward models that are small enough and portable enough to actually ship.

This is where the open-weight history becomes a technical asset rather than a cultural one. An organization that has spent years releasing models people download and run locally has, by necessity, built muscle around distribution, quantization tolerance, and inference outside its own infrastructure. A company whose entire stack assumes its own serving environment has to retrofit all of that. Nous starts from a position where portability was the default rather than the exception.

The architectural tension is that agents want the opposite of portability. Agent systems benefit from centralized orchestration, shared tool registries, persistent memory stores, and tight telemetry. Reconciling “runs inside your walls” with “coordinates a long multi-step process reliably” is a genuinely hard design problem, and it is the one I would most want to see Nous describe in detail.

What I will be watching

The open-weight community will ask whether Hermes stays open. That is a fair question but not the one I find most informative. Plenty of companies maintain an open base and a commercial layer on top without the base degrading.

The more diagnostic signal is whether enterprise requirements start flowing backward into the model architecture itself. Enterprise agents need calibrated uncertainty, reliable tool-call formatting, graceful refusal when a step is underspecified, and consistent behavior across long contexts. Those are training objectives, not wrapper features. If future Hermes releases show measurable improvement on those properties, the enterprise push will have made the open models better. If the open releases drift toward general-purpose chat while the commercial product gets the agentic reliability work, that tells a different story.

Either way, $158 million buys enough runway to find out which kind of company this becomes. The capital was raised on a thesis about agents, and agents are judged on whether they finish the job.

🕒 Published: October 8, 2026