AI · November 18, 2026

Also published in Deutsch, Español, Norsk

Mirendil raises one billion dollars in Series A funding

A yellow robotic arm with black joints and a mechanical hand
Enchanted Tools / Unsplash

Mirendil, a software provider that accelerates artificial intelligence development, has raised one billion dollars in its initial funding round. The startup announced on Thursday that Kleiner Perkins and Clearlake Capital co-led the Series A investment. More than six other investors participated in the round, including Google.

The company was founded last year by former researchers from Google DeepMind and Anthropic. Behnam Neyshabur serves as the chief executive officer and is a co-inventor of SAM, a popular algorithm that helps artificial intelligence models perform tasks they were not specifically trained for. Harsh Mehta, the chief technology officer, focused on pretraining projects during his tenure at Google DeepMind and Anthropic. Pretraining is the term used for the initial stage of the artificial intelligence development process.

Mirendil's first product, Lor, is a platform that automates various tasks during the model training phase. The company states that the software operates by dividing work into smaller parts and assigning them to hundreds of artificial intelligence agents. Developing large language models involves extensive trial and error. Researchers create multiple versions of an algorithm that differ only slightly and compare them to find the most efficient version. According to Mirendil, Lor can execute these workflows automatically. If an error occurs during evaluation, the platform corrects it automatically.

Earlier this year, Mirendil assigned a single researcher to build a coding model capable of competing with leading open source large language models. Such projects are typically carried out by teams of experts. Mirendil states that its automation technology allowed the researcher to complete the model without help from colleagues. The project cost ten million dollars and took approximately three and a half months. The company's agents wrote more than ninety nine percent of the code that powered the model, named Mirendil-Base. These agents set up the infrastructure, found suitable training algorithms, and gathered training data from the web. The agents then improved both the training dataset and the architecture of Mirendil-Base to enhance the quality of the results.

This large language model serves as the foundation for a specialized coding model named Fern, which the company launched last month. Mirendil data indicates that Fern is designed to automatically handle programming challenges often encountered by artificial intelligence laboratories. Fern can process ten million tokens per prompt, which is ten times more than Claude Opus 5.5. This capacity is equivalent to approximately half a million lines of code. Fern was developed from a version of Mirendil-Base with between one hundred billion and five hundred billion parameters. At the end of last month, the company demonstrated two upgraded versions of Mirendil-Base with parameters ranging from five hundred billion to ten trillion. Mirendil plans to use these algorithms to develop more capable large language models in the future.