A billion dollars for Mirendil’s AI research tools

A $1 billion Series A backs Mirendil’s coding models and Lor, an agent platform still in private beta.

A $1 billion funding round backs Mirendil’s artificial intelligence research tools. The Series A is led by Kleiner Perkins and Clearlake Capital, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC, and S32 among the participants. The lab develops both its own models and a platform where agents write code, run experiments, and examine their results.

Co-founded by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, the company had already raised $200 million less than six months earlier, according to the Wall Street Journal. The newspaper reports that the new round values the company at $5 billion, citing people familiar with the matter. That figure does not appear in Mirendil’s announcement.

The company’s thesis is to assign agents a growing share of the engineering work surrounding research: preparing data, coding experiments, monitoring training runs, and investigating failures. The funding is intended to support tools that make these capabilities accessible to scientists and engineers outside the lab. The initial published results primarily concern model development and research code; they do not yet demonstrate general automation of scientific discovery.

The central experiment is the pretraining of Mirendil-Base. According to the lab, one researcher completed it in three and a half months using less than $10 million in compute. That budget covers preliminary small-scale experiments, data processing, agents’ token consumption, and the final training run. The provider says agents wrote more than 99% of the code, while the researcher selected directions, assessed results, and intervened when necessary.

The project reportedly started from an empty repository, using libraries such as PyTorch and JAX. Training data reportedly came from public sources, without distillation or synthetic data. The announced cost applies to Mirendil-Base, the smallest model in the family, rather than describing the budget for the other versions.

The published technical comparison measures code prediction in bits per byte, rather than the percentage of coding tasks solved. It relies on evaluations conducted by the lab.

The Fern 0.1 model results from a short post-training stage applied to Mirendil-Base. Its announced context window reaches 10 million tokens, accepting text and code. The provider targets reading lengthy execution logs, repositories, and agent traces to retain more information about ongoing work.

That stated capacity does not, by itself, establish reasoning quality at that length. The presentation mentions an initial context-access test using RULER and says more comprehensive evaluations will follow. Model size is disclosed only as a range, between 100 billion and 500 billion parameters. Its page still lists access as coming soon, with an early-access request option.

The Lor platform is in private beta. It can use Mirendil’s models and third-party models to distribute work across dozens or hundreds of agents. Its architecture separates the cloud-based agent from the machine executing its commands. Tasks can continue on a remote computer or another machine that remains powered on when the user closes their laptop.

The service also provides compute: agents request nodes, configure them, run their work, and release them afterward. Examples include reproducing a scientific experiment across multiple GPUs, running training variations, or examining a software repository. Users can monitor and stop agents, which operate only on machines added to the system.

The technical post specifies that Mirendil-Base did not build itself and does not present the experiment as fully autonomous recursive self-improvement. The researcher continues to choose which directions to explore and which results to trust.