AI lab Mirendil has raised a $1 billion Series A led by Kleiner Perkins and Clearlake Capital, one of the largest rounds in AI this year. The company did not disclose a valuation.
Announced Thursday, November 12, the round comes less than five months after Mirendil emerged from stealth. Participants include Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32. The lab has now raised about $1.2 billion, including a $200 million seed round in June.
Mirendil's pitch: AI agents can now do the hands-on work of building frontier models, so a small team can do what used to take an entire institution. The money bets that the experiment the company published three weeks earlier is a beginning, not a stunt.
The $10 million experiment behind the raise
On October 21, Mirendil published results from a research run in which a single researcher, directing AI agents, pretrained a coding model called Mirendil-Base from scratch in three and a half months, for under $10 million in compute — including the agents' token spending and data processing.
Agents wrote more than 99 percent of the code behind the run, ran experiments and diagnosed failures, while the researcher decided what to trust. Mirendil is careful to say the model did not build itself: the work depended on an experienced researcher directing the agents and judging their output.
The company says Mirendil-Base sits in the 100-billion to 500-billion parameter range. Two larger versions were still training when the report was published.
Mirendil then post-trained the base model into Fern 0.1, its first model, built for its own researchers' agentic work: debugging training runs, reading agent traces and auditing training data. Fern claims a 10-million-token context window.
From internal tool to product
The product that turns that internal workflow into something others can use is Lor, introduced on October 28. It coordinates hundreds of agents that write code, run experiments and iterate for days or weeks, and it remains in an invitation-only private beta, free during the beta period.
Mirendil was founded in June by 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. The lab announced a multi-year Google Cloud partnership in August, running its training and inference on Google Cloud's AI Hypercomputer.
Kleiner Perkins led both of Mirendil's rounds. The June seed was led by Andreessen Horowitz and Kleiner Perkins at a reported valuation of about $1 billion, with Nvidia investing afterwards.
No valuation, and a test the company set itself
What the announcement does not say is what Mirendil is worth now. Bloomberg reported in September that it 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 the final terms match that report.
There are caveats. The sub-$10 million figure covers only its smallest model, and the performance evidence so far is a company-run test: in Mirendil's bits-per-byte comparison on seven external code repositories, Mirendil-Base ranked second to Kimi-K3. No independent evaluation of any Mirendil model has been published.
Still, the scale of the round — the biggest US startup financing of the week, ahead of LangChain's $300 million Series C — shows investors are willing to fund the idea that the most expensive work in AI, building the models themselves, can be handed to agents. Whether a small team with a billion dollars can keep pace with labs spending far more is the question the money must answer.
