Mirendil raises $1B Series A to build AI that does AI research

A lab that wants AI to do AI research just raised $1 billion to prove it. Mirendil, founded by former Anthropic researchers, announced a $1 billion Series A on November 12, led by Kleiner Perkins and Clearlake Capital. It is one of the largest early-stage rounds in AI this year, and it is a bet on a very specific claim: that agents can now do the hands-on work of building frontier models, and that a small team can therefore do what used to take an institution.

The pitch is not a slide deck. Three weeks before the raise, Mirendil published results from an experiment in which a single researcher, directing AI agents, pretrained a coding model from scratch in three and a half months for less than $10 million in compute. The round is investors deciding that result is a beginning, not a stunt.

The round: a billion dollars, no valuation#

The financing landed less than five months after Mirendil emerged from stealth. At its launch on June 24, the company disclosed a $200 million seed round led by Andreessen Horowitz and Kleiner Perkins, at a reported valuation of about $1 billion, with Nvidia investing afterwards. Kleiner Perkins has now led both rounds, and Clearlake Capital joins as co-lead of the Series A.

The investor list also includes Google, which is both a backer and a supplier: Mirendil announced a multi-year Google Cloud partnership in August, putting its training and inference workloads on Google Cloud’s AI Hypercomputer, drawing on TPUs and Nvidia infrastructure. The founding team numbers about 20 researchers and engineers drawn from Anthropic, xAI, Google DeepMind and OpenAI, led by chief executive Behnam Neyshabur and chief technology officer Harsh Mehta, both formerly of Anthropic.

What the announcement does not say is what the company is worth now. Bloomberg reported in September that Mirendil was in talks to raise up to $1 billion at a $5 billion valuation, but the Series A release names no valuation, and no source has confirmed that the final terms match the earlier report.

The claim behind the money#

The idea is spelled out in a research report Mirendil published on October 21. One researcher, working with AI agents, pretrained Mirendil-Base, a coding model in the 100-billion-to-500-billion parameter range, from scratch in 3.5 months. The all-in cost, Mirendil says, was under $10 million in compute, covering research and small-scale experiments, the final training run, agent token spending and data processing.

According to the report, agents wrote more than 99% of the code, ran the experiments and diagnosed failures, while the researcher chose what to try, judged what to trust and intervened when needed. The training used publicly available code, web pages and public datasets, with no purchased or licensed data and no distillation or synthetic data. Two larger versions, Mirendil-Base-Large and Mirendil-Base-XL, were still training when the report was published.

That is the result the $1 billion is really funding. If a single researcher plus agents can pretrain a competitive coding model for under $10 million, the economics of frontier research change: the bottleneck stops being headcount and compute budgets and becomes the quality of the agents doing the work.

Fern and Lor: the product stack#

Mirendil then post-trained the base model into Fern 0.1, its first model, built for the kind of high-volume agentic work its own researchers do: debugging training runs, reading complete agent traces and auditing training data. Fern holds up to 10 million tokens of context, roughly 23 megabytes or half a million lines of code, enough to keep days of an agent’s work in view at once. Access is listed as coming soon.

Lor is the platform that turns the internal workflow into a product. Introduced on October 28, it is a cloud-based system whose agents write code, run experiments, provision compute and operate on machines the user selects. An agent can split a job across dozens of agents, with hundreds running across projects, and the company says they can keep going for days or weeks and unblock themselves when something breaks. Lor is in private beta, invitation-only and free during the beta, on macOS, Linux and the web, with Windows, iOS and Android listed as coming soon.

What the money is for — and what to be skeptical about#

The pitch, in Mirendil’s own framing, is that one person should be able to take on what used to need a team, and a team what used to need an institution. The Series A buys the compute and the hiring to test whether that scales: bigger models, a wider Lor beta and the infrastructure to run research loops that last days instead of hours.

The honest caveats are substantial. None of the headline numbers has been independently checked: there is no outside evaluation of Mirendil-Base or Fern, no public benchmark suite behind the “competitive coding model” description, and the larger models were still training when the report came out. The announcement says nothing about safety checks or an access policy for agents that provision their own compute. And the one-researcher result, however striking, is a single experiment run by a highly experienced researcher directing the agents — the open question is whether the same loop works in the hands of ordinary scientists and engineers, which is the entire point of Lor.

What to watch#

Three things will tell us whether the bet pays off. First, whether Lor’s private beta widens and what outside users actually manage to build with it — the product has to work for scientists who are not Mirendil researchers. Second, independent evaluations of Mirendil-Base and Fern, which would turn a company-reported result into a verified one. And third, results from the larger Base-Large and Base-XL models, which will show whether the one-researcher method scales or was a one-off. A billion dollars does not verify the claim; it finances the attempt to prove it at scale.

Sources: Mirendil’s Series A announcement; Unite.AI; TechTimes; Securities.io.