Picture a single researcher, working alone in a small San Francisco office, training a frontier AI model from scratch. No army of engineers. No billion-dollar compute cluster booked months in advance. Just one person, and a swarm of AI agents doing the heavy lifting beside them.
That is the scene Mirendil says is already real. And this week, some of Silicon Valley’s most powerful investors decided it is worth a billion dollars.
The startup announced on November 12 that it has raised a $1 billion Series A, led by Kleiner Perkins and Clearlake Capital. The round comes barely five months after Mirendil emerged from stealth with a $200 million seed round, bringing its total disclosed funding to roughly $1.2 billion in less than a year.
For a company that is barely a year old, that is an extraordinary vote of confidence. It is also a bet on a radical idea: that AI research itself can be handed over to AI agents.
The labs of the future, in this vision, will be run by a handful of people directing hundreds of tireless digital researchers.
Inside the Mirendil Series A
The Series A was led by Kleiner Perkins and Clearlake Capital, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 joining in, alongside other new and existing investors. Nvidia, which backed the seed round, was among the earlier supporters.
Mirendil did not disclose a valuation for the new round. Its seed round in June was struck at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins. A jobs newsletter has put the new valuation at $5 billion, but that figure could not be confirmed, and Mirendil has not commented on it.
What is clear is the speed. Five months separated the seed announcement from a Series A five times its size. In a year when investors have grown more cautious about AI valuations, a billion-dollar cheque for a company with a product still in private beta is a statement.
Who Is Behind Mirendil?
Mirendil was founded in early 2026 by Behnam Neyshabur and Harsh Mehta, two researchers who left Anthropic in December 2025. They are joined by Shayan Salehian, formerly of xAI, and Tara Rezaei, who previously interned at OpenAI.
The founding team of around 20 researchers and engineers was drawn from Anthropic, xAI, Google DeepMind and OpenAI. It is, in other words, a team built from the very labs whose frontier models they now hope to help smaller teams reproduce.
Their pitch is simple. Frontier AI research has become so expensive and labour-intensive that only a handful of companies can do it.
Mirendil wants to change that by building agents that do the hands-on work of research, so that one person can take on a project that once needed a whole team, and a small team can attempt work that once needed an entire institution.
One Researcher, One Model, 3.5 Months
The claim that underpins the fundraising was published on October 21, in Mirendil’s first research disclosure. According to the company, a single researcher working with AI agents pretrained a coding model called Mirendil-Base from scratch in 3.5 months, using less than $10 million of compute.
If true, that is a striking result. Pretraining a frontier-class model is normally a months-long effort involving large teams and budgets measured in the tens or hundreds of millions of dollars. Mirendil says its agents handled the work that would usually fall to those teams.
Mirendil-Base sits in the 100 billion to 500 billion parameter range. Larger versions, Mirendil-Base-Large and Mirendil-Base-XL, were still in training when the report was published. The company then post-trained the base model into Fern 0.1, its first model, which it says can hold up to 10 million tokens of context.
That context window matters. A 10-million-token window means Fern can read an entire codebase, a full set of agent traces, or a mountain of training data in a single pass. Mirendil describes it as a system built for high-volume agentic work, from debugging training runs to auditing data.
Meet Lor, the Platform Running Hundreds of Agents
The product at the centre of Mirendil’s plan is called Lor. Introduced on October 28, it is a cloud-based platform that lets users delegate research tasks to hundreds of AI agents working in parallel.
According to Mirendil, Lor’s agents write code, run experiments, provision compute and operate on machines chosen by the user, iterating for days or even weeks at a time. The company says it uses Lor every day for its own research.
For now, Lor is in an invitation-only private beta and free during that period. It runs on macOS, Linux and the web, with Windows, iOS and Android versions listed as coming soon. Fern access is also listed as coming soon.
The company has not disclosed how many external users Lor has, or how much work they have completed with it. That is worth keeping in mind: so far, the most impressive demonstration of the system is Mirendil’s own.
Why the Mirendil Series A Matters
The funding lands at a moment when the AI industry is asking hard questions about who gets to build frontier models. Training costs have soared, compute is scarce, and the gap between the biggest labs and everyone else keeps widening.
Mirendil’s answer is to shrink the team, not the ambition. If agents can do the grinding work of pretraining, debugging and experimentation, then a university lab, a startup or even a single researcher could attempt projects that today belong only to companies worth hundreds of billions of dollars.
That is also why Google is both an investor and a partner. Mirendil announced a multi-year Google Cloud partnership in August, covering training, inference and AI research workloads. The cloud capacity that makes its experiments possible comes, in part, from the very giants it hopes to democratise against.
The timing is notable in another way. The same week Mirendil announced its round, four nonprofits sued OpenAI, Google, Anthropic, Meta and SpaceX’s AI division, asking a California court to freeze frontier AI development until the labs can prove their systems are under control.
The industry’s power is being questioned from the outside, even as Mirendil tries to spread it more widely from the inside.
The Scepticism Is Fair
None of Mirendil’s core claims has been independently verified. There is no public evaluation of Mirendil-Base or Fern, no independent timing of the 3.5-month pretraining run, and no published benchmark showing how the models compare with the frontier systems they are meant to rival.
The company has also disclosed no valuation, no revenue, and no user numbers for Lor. The $1 billion is, for now, a bet on a story and a team rather than a proven product. Kleiner Perkins and Clearlake are betting that the story holds up.
That scepticism does not make the idea uninteresting. The pattern Mirendil describes, agents doing the routine work of research while humans set the direction, is one the whole industry is moving toward.
Google’s Gemini 4 Hydrogen is built for coding agents, and Anthropic’s Claude office suite is aimed at the same agentic future. Mirendil is simply pushing the logic further, into the lab itself.
What Happens Next
The billion dollars buys Mirendil time and compute. Expect Lor to open more widely, Fern to become available, and the larger Mirendil-Base models to finish training. The real test will be whether outside researchers, not just Mirendil’s own team, can use the platform to do work they could not do before.
If they can, the economics of AI research change. A $10 million pretraining run done by one person is not just cheaper than the alternative, it is a different world, one in which the frontier is no longer reserved for the richest few.
If they cannot, Mirendil will be remembered as one of the most expensive bets of the AI boom on an idea that arrived too early. Either way, the next year will tell. The agents, as Mirendil likes to say, are already working.
According to Unite.AI, the round was announced on November 12, 2026.
