Open-source models usually spread like mushrooms after rain: fast, everywhere, and nobody can tell you exactly who planted them. Nous Research just did the unusual thing and put a price tag on the forest floor. The three-year-old startup behind the open-source Hermes agent confirmed a $1.5 billion valuation on the back of a $90 million Series B, and is now pointing all of that at a product called “Hermes for Businesses.”
I build bots for a living, so my first reaction wasn’t about the money. It was about the 24 million clones.
What 24 million clones actually tells you
Hermes has been cloned more than 24 million times and accounts for roughly 2.5% of global AI token usage. Those two numbers together are more interesting than either one alone. Clone counts measure curiosity. Token usage measures people actually shipping something.
Plenty of open-weight projects rack up downloads and then live out their days in a folder named experiments/. A 2.5% share of worldwide token traffic means Hermes is in production loops somewhere, grinding through real requests on somebody’s budget. For anyone who has tried to convince a team to standardize on an open model instead of a hosted API, that figure is the most useful piece of ammunition to come out of this announcement.
It also sets up the obvious question: if your model is already everywhere, how do you make money from it? Nous’s answer appears to be the one most open-source companies land on eventually. Keep the weights open, sell the operational layer on top.
The money, briefly
The round was led by Robot Ventures, with backing that includes Nvidia, Union Square Ventures, and Samsung. That takes total funding to $158 million. The Wall Street Journal reported the deal on October 7, 2026, after TechCrunch had flagged a round of at least $75 million forming back in July.
Revenue is where it gets concrete. Nous was at roughly $36 million in annualized revenue by mid-September 2026 and expects to pass $100 million by year-end. That’s a steep curve to promise out loud, and it tells you the enterprise product isn’t a speculative side quest. Someone has already signed contracts.
Why “multi-step workflows” is the whole ballgame
Hermes for Businesses lets companies deploy customized agents that handle multi-step workflows. That phrasing is doing a lot of quiet work, and if you’ve ever built an agent that has to survive more than three tool calls, you know why.
Single-turn is easy. Prompt in, answer out, done. Multi-step is where the wheels come off:
- State drifts. Step four forgets what step two decided, and nothing in the output tells you it happened.
- Failures compound. A 95% success rate per step is a coin flip by step twelve.
- Side effects are permanent. A hallucinated summary is annoying. A hallucinated API call that updates a customer record is a Monday morning.
- Debugging is archaeology. Without proper traces, you’re reconstructing a crime scene from a single stack trace.
Whoever solves the boring parts of this — retries, checkpoints, observability, clean rollback semantics — wins enterprise agent work. Not whoever has the cleverest reasoning benchmark. If Nous is charging real money for multi-step reliability, that’s a bet on plumbing, and it’s the right bet.
What I’d check before committing
The announcement is light on technical specifics, which is normal at this stage. So here’s my own checklist, the same one I’d apply to any agent platform pitching my team:
- Where does the model run? Open weights mean you can self-host. Does the business tier preserve that, or does customization quietly route you through someone else’s endpoint?
- What does a trace look like? I want per-step inputs, outputs, tool calls, and timing. If I can’t replay a failed run, I can’t support it.
- How are tools defined? Portable schemas mean I can migrate. Proprietary tool formats mean I’m renting my own integration work back from the vendor.
- What happens mid-failure? Does a halted workflow leave half-applied changes, or is there a compensation path?
- How far does customization go? Prompt templates and fine-tuning are very different products at very different prices.
The pattern worth watching
The useful signal here isn’t the valuation. It’s the shape of the business: give away the model, charge for making it dependable at scale. Hermes earned its distribution the hard way, through 24 million clones and a measurable slice of global token traffic, and now the company is monetizing the gap between “this model works” and “this agent survives Tuesday in production.”
That gap is where most of us spend our actual working hours. Somebody finally decided it was a product.
🕒 Published: