Think of an open-weights model as a food truck. It parks wherever you want, you can pop the hood, and nobody charges you a cover fee. Enterprises, meanwhile, have been eating in hotel restaurants — predictable menus, locked kitchens, a bill that arrives per seat. Nous Research just announced it is parking the truck in the hotel lobby.
The company confirmed a $1.5 billion valuation alongside a $90 million Series B led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and 1789 Capital participating. That takes total funding to $158 million. The money funds a push into the enterprise sector under a product called “Hermes for Businesses,” which lets companies deploy customized agents that handle multi-step workflows while keeping data private and secure.
For those of us who build bots for a living, that product description is the part worth reading twice.
Why the privacy line matters more than the valuation
I have lost count of how many agent projects died in a procurement meeting. Not because the agent did not work. Because someone in legal asked where the prompts go, and the answer involved a third-party endpoint in another jurisdiction. You can build the cleanest tool-calling loop in the world and still watch it stall out on a data residency question.
Nous has been shipping Hermes as an open model, and open weights solve that problem at the architecture level rather than the contract level. If the model runs inside your own boundary, the data question becomes an infrastructure question, which is a question enterprise IT teams already know how to answer. The pitch for Hermes for Businesses sits exactly there: customized agents, your workflows, your privacy posture.
What is not public yet is the shape of the offering. Managed deployment? A control plane over self-hosted weights? Fine-tuning as a service? Those details decide whether this is interesting to a platform team or interesting to a VP who wants a vendor to call. I would not guess, and you should not either until the docs land.
Multi-step workflows are where agents actually break
“Multi-step workflows” is doing a lot of quiet work in that announcement. Anyone who has shipped an agent knows the single-call demo is easy and step seven is where everything falls apart. State drifts. A tool returns a malformed payload and the model confidently narrates around it. Retries stack into a loop that burns tokens for ninety seconds and produces a summary of its own confusion.
So when I evaluate something like this, the questions I care about are not benchmark scores:
- How is intermediate state stored, and can I inspect it mid-run?
- What happens on a tool failure — does the run halt, retry, or improvise?
- Can I pin a model version so a silent upgrade does not change my agent’s behavior overnight?
- Is there a trace I can hand to an auditor, or just a chat log?
- What does the fine-tuning path look like for a company-specific tool schema?
An open-weights foundation gives you a better shot at answering most of those yourself, which is the structural advantage here. You are not filing a support ticket to find out why your agent looped.
Reading the investor list
The cap table is an odd mix and tells you something. Nvidia and Samsung are strategic hardware money, the kind that shows up when a company plans to run a lot of compute, possibly on customer premises. Union Square Ventures and Menlo Ventures are traditional software bets. 1789 Capital, where Donald Trump Jr. is a partner, brings a political dimension to a company whose core argument is that open models are the safer enterprise choice. Open-weights AI has become a policy conversation as much as a technical one, and that mix of backers reads like a company positioning for both.
Reporting on the round has been inconsistent, too. Earlier coverage described a raise of at least $75 million; the confirmed figure is $90 million. Revenue figures circulating in secondary coverage have not been confirmed by the company, so treat them as unverified until they are.
What I would do with this
Nothing yet, honestly. I would read the Hermes for Businesses documentation the day it is public, then run the same evaluation I run on every agent framework: a five-step workflow with one tool deliberately returning garbage, and see what the trace looks like.
But the direction is one I have wanted for a while. The agent stack has been splitting into “powerful but hosted elsewhere” and “self-hosted but you assemble everything.” A credible middle option — open weights with enterprise tooling around them — removes the reason most of my agent prototypes never made it past review. That is worth a lot more to a builder than a valuation headline.
🕒 Published: