One researcher, working with a fleet of AI agents and less than $10 million in compute, built Mirendil’s first foundation model, the company says. Investors have now handed the startup $1 billion to see how far that approach can go.
The San Francisco lab said Thursday it raised a $1 billion Series A led by Kleiner Perkins and Clearlake Capital, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also taking part. The round is a wager that agents can take over much of the hands-on engineering behind frontier AI. Mirendil says pretraining a competitive model has taken dozens of specialists at open-model efforts and hundreds at frontier labs, and that one researcher did it with agents. The company wrote in its announcement that “when agents do the hands-on work, the size of your team stops limiting what you can build.”
Mirendil didn’t disclose a valuation. A talent newsletter published by Andreessen Horowitz, which co-led the company’s seed round, put it at $5 billion, five times the seed valuation. Bloomberg reported on Sept. 22, citing people familiar with the matter, that Mirendil was in talks to raise up to $1 billion at a $5 billion valuation including the new money. The step-up has been fast. Mirendil came out of stealth on June 24 with a $200 million seed round led by Andreessen Horowitz and Kleiner Perkins, followed by an investment from Nvidia. SiliconANGLE reported the seed valued it at $1 billion. In August, it signed a Google Cloud partnership worth more than $100 million for a mix of TPUs and Nvidia GPUs, the a16z newsletter said.
The founders come from the labs Mirendil now wants to compete with. Chief Executive Behnam Neyshabur co-led Google DeepMind’s Blueshift team, which worked on reasoning and math for Gemini. Chief Technology Officer Harsh Mehta built the first version of Anthropic’s internal AI research platform. The two joined Anthropic in late 2024 and left together in December 2025 to start the company, according to Dealroom. The a16z post describes a team of about 20 researchers and engineers.
One researcher, 3.5 months
The headline result is Mirendil-Base, a coding-focused model in the 100 billion to 500 billion parameter range. In an Oct. 21 technical post, Mirendil said one researcher built it from an empty repository in 3.5 months. Agents wrote more than 99% of the code, ran the experiments and diagnosed failures, while the researcher decided what to try and what to trust. The company said the sub-$10 million figure is all-in, covering the agents’ token spend and data processing, and that it used only publicly available data, with no distillation and no licensed datasets. Two larger versions are in training, one with 500 billion to 2 trillion parameters and another with 2 trillion to 10 trillion.
Mirendil then post-trained the base model into Fern, which can hold up to 10 million tokens of context. That is roughly half a million lines of code, or days of an agent’s work, in view at once. The company is pitching it at the routine work of model building, such as debugging a training run from its logs or scanning a pretraining shard for duplicates and benchmark contamination. Mirendil is taking early-access requests for Fern.
The commercial product is Lor, an agent platform now in invite-only private beta. Its agents write code, run experiments and iterate for days, provision their own GPU nodes and release them when done, and can split a job across hundreds of parallel workers. Mirendil says it is building Lor “to do the work of an entire AI lab.” Lor runs Anthropic’s and OpenAI’s frontier models alongside open-source models and Mirendil’s own. In one example the company gave, 102 agents built training environments from 100 PyTorch pull requests, averaging eight minutes each. In another, agents found security bugs in LiteLLM overnight and fixed them by morning.
The round puts Mirendil among the best-funded of the new research labs, a group Dealroom said includes Safe Superintelligence, Thinking Machines Lab and Recursive Superintelligence. It also lands about a week before Anthropic, where much of the founding team worked, is set to list on Nasdaq after setting its IPO price range. Lor is free during the private beta. Mirendil has not said when it will open beyond its waitlist, what it will cost afterward or when the larger Mirendil-Base models will finish training.
WebProNews is an iEntry Publication