Mirendil Raises $1B Series A to Scale One-Researcher AI Development

Mirendil has raised $1 billion in a Series A led by Kleiner Perkins and Clearlake Capital, the company announced on November 12, 2026, to make frontier AI research and development accessible to more scientists and engineers. That figure lands barely five months after the company's June 24, 2026 launch and on top of a $200 million seed round led by Andreessen Horowitz and Kleiner Perkins, with follow-on investment from NVIDIA and others. A $1 billion Series A for a company this young is rare, and it rests on an internal result that has not yet been independently reproduced.

Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also participated in the round, alongside other new and existing investors. Mirendil is building models and agent software around a single organizing goal: enabling one person to take on work that previously required a team, and a team to pursue projects that previously required an institution. The claim is ambitious. The evidence so far comes from the company's own published report.

What did the one-researcher experiment actually show?

In an October 21, 2026 report, Mirendil said one researcher working with AI agents pretrained Mirendil-Base, a coding model, from scratch in 3.5 months with less than $10 million in compute. The all-in figure covers research and small-scale experiments, the final training run, agent token spending and data processing. A frontier-class model trained for under $10 million would undercut prevailing frontier training budgets by orders of magnitude — if the result holds up outside the company's own setup.

Mirendil reports that agents wrote more than 99% of the code, ran experiments and diagnosed failures. The researcher chose what to try and trust and intervened when needed. That division of labor is the core of the company's accessibility pitch: the human supplies judgment, the agents supply labor.

The experiment drew only on publicly available code, web pages and public datasets, without purchasing or licensing additional data and without distillation or synthetic data. That data provenance matters because it removes the most common legal and cost objections to replicating the approach. Mirendil-Base sits in the 100 billion to 500 billion parameter range. The larger Mirendil-Base-Large and Mirendil-Base-XL models were still in training when the report was published, so the flagship scale results remain unseen.

Takeaway: one internal training run, self-reported, is the entire foundation under the round.

What are Fern and Lor?

Mirendil subsequently post-trained Mirendil-Base into Fern 0.1, its first model. The Fern model page describes a system built for high-volume agentic work: debugging training runs, reading complete agent traces and auditing training data. The architecture, training and serving support a context window of up to 10 million tokens — roughly 23 MB or half a million lines of code — which the company says lets the model hold days of an agent's work at once. That context length is the technical hook; a model that can ingest an entire training-run history in one pass changes how debugging and auditing are done. Access is listed as coming soon, which means the claim is not yet publicly testable.

Lor is the platform for delegating that work. Mirendil introduced Lor on October 28, 2026 as a cloud-based system whose agents can write code, run experiments, provision compute and operate on machines selected by the user. An agent can divide a job across dozens of agents, with hundreds running across projects at once. The company says agents can continue for days or weeks and unblock themselves when something breaks. Autonomy at that timescale is exactly where current agents tend to fail, so the durability claim deserves scrutiny once outside users get hands on the product.

Lor is available by invitation in private beta and is free during the beta on macOS, Linux and the web. Windows, iOS and Android are listed as coming soon. Users can follow each agent's progress, stop agents and remove machines from the system — control surfaces that suggest Mirendil expects users to want a kill switch for long-running autonomous work.

Takeaway: Fern supplies the model, Lor supplies the workforce, and both remain behind invite walls.

Who is building it, and on whose compute?

At its June 24, 2026 launch, Mirendil described a founding team of 20 researchers and engineers drawn from Anthropic, xAI, Google DeepMind and OpenAI. That lineage puts the startup in direct competition with the labs its founders left, and it explains why investors priced the round at Series A size with seed-stage history.

The compute side is already locked in. Mirendil announced a multi-year Google Cloud partnership on August 6, 2026 that puts training, inference and AI research workloads on Google Cloud's AI Hypercomputer, drawing on Google Cloud TPUs and full-stack NVIDIA AI infrastructure. The arrangement includes managed training clusters co-designed across compute, storage, networking and control planes to support larger research loops. Google's participation in the Series A now looks less like a neutral bet and more like an infrastructure customer taking equity in its tenant.

Takeaway: the team has frontier-lab credentials and committed infrastructure, but no externally validated product yet.

Why does this round matter?

The $1 billion Series A connects three parts of Mirendil's plan: agent-assisted model development, a platform for delegating research work, and cloud infrastructure for larger experiments. Each part exists today only as a company demonstration or a private beta. The stakes extend beyond one startup. If one researcher can genuinely train a competitive model for under $10 million, the economics of frontier AI shift from capital concentration toward distribution of capability — the stated mission behind the round. If the result does not transfer, the round becomes another bet on a team rather than a method.

The company has been transparent about its methods on paper: public data only, no distillation, full cost accounting in the October report. But every load-bearing number — the 99% agent-written code, the 3.5-month timeline, the sub-$10 million total — is self-reported. No third-party audit or external replication appears in the source material.

The central test, as the company frames it, is whether the one-researcher result can become a repeatable workflow for outside scientists and engineers. Private beta invitations and a "coming soon" label on Fern are the first gates; broad public access and an independent replication of the training result would be the real verdict on whether the accessibility mission is marketing or method.