Mirendil, a young AI lab, said on Thursday, Nov. 12, that it has raised $1 billion. Kleiner Perkins and Clearlake Capital led the round, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 among the other investors. Mirendil's two founders worked on Google's Gemini models and then at Anthropic, according to Kleiner Perkins, which has backed the company from the start.
Mirendil is betting that AI agents, AI programs that can carry out tasks on their own such as writing code and running experiments, can now do most of the hands-on work of making a new AI model. Today, it says, a lab that wants to use AI in drug discovery, chemistry, biology or robotics must also become a cutting-edge AI lab itself, which is expensive and needs skills found in only a few labs. Mirendil says it wants to put the work of building advanced AI within reach of every scientist and engineer.
Its main evidence is Mirendil-Base, a coding model that it says one experienced researcher, working with AI agents, built from scratch in three and a half months, as it described on Oct. 21. The work covered pretraining, the first, broad stage of training, when a model learns from a huge collection of text. Mirendil says agents wrote more than 99% of the code, while the researcher decided what to try and what to trust, and stepped in when needed. Mirendil puts the full cost, including the cost of running the agents, at less than $10 million in computing power. That figure covers only Mirendil-Base, the smallest of three models; two larger ones were still in training as of Oct. 21. Mirendil says that for most of the past decade, building such a model took a team: dozens of specialists at groups that release their models for anyone to download, and hundreds at the biggest labs.
Neither of the two products Mirendil announced in late October is open to everyone yet. Fern, a version of Mirendil-Base given extra training for coding jobs, can hold up to 10 million tokens at once. Tokens are the small chunks of text, often parts of words, that AI models read and write. Mirendil says 10 million of them is about half a million lines of code. Fern is listed as coming soon, and Mirendil says it is opening early access to a select few teams and researchers. Lor, an app Mirendil announced on Oct. 28, where it says hundreds of agents can write code and run experiments for days, is open only to invited testers and is free for now. It also runs other companies' models, including Anthropic's Claude, OpenAI's GPT-6 and Google's Gemini.
The test results so far are Mirendil's own. On the company's table of five coding tests, Fern scores below OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 on every one. It also trails two models anyone can download, Kimi K3 from Moonshot AI and GLM-5.3 from Z.ai, on all five, though Mirendil says this first version had only a short round of extra training. On a test of predicting code from seven outside projects published after the models were trained, Mirendil-Base came second of five models, behind Kimi K3. On Oct. 27, Mirendil corrected two charts on Fern's page about how well it handles long inputs, saying the new versions more accurately reflect its training results. It says fuller results on long inputs are still to come.
Mirendil has not said what the new round values the company at. An earlier round, $200 million in June, valued it at $1 billion, according to a news report. It has not said when Fern or Lor will open to everyone, or how it will spend the money, beyond a broad aim and a call for new hires.