Tuesday, November 17, 2026Tue, Nov 17 Updated

Mergers, acquisitions and startup deals, and what they mean.

Mirendil Raises $1 Billion Series A to Automate AI Model Training

AI startup Mirendil raised $1 billion in a Series A led by Kleiner Perkins and Clearlake Capital, with Google joining. What its Lor platform does and who runs it.

High-tech server rack in a secure data center with network cables and hardware components.
Photo: Sergei Starostin / Pexels

Mirendil Inc., a startup that builds software to speed up the development of artificial intelligence models, has raised $1 billion in a Series A round, the company said on Thursday. Kleiner Perkins and Clearlake Capital led the round together, and more than a half-dozen other investors took part, including Google, according to SiliconANGLE.

A Series A is usually a company’s first major institutional round. A raise of this size for a company founded only last year gives Mirendil a large budget for computing and staff as it tries to automate much of the work AI labs now do by hand. The company’s valuation was not disclosed.

Who founded Mirendil

Two researchers who previously worked at Google DeepMind and Anthropic started the company last year. Chief Executive Officer Behnam Neyshabur co-invented SAM, a widely used algorithm that helps AI models perform better on tasks they were not trained for directly. Chief Technology Officer Harsh Mehta worked on pretraining at both labs. Pretraining is the first stage of building an AI model, when it learns from large amounts of data.

What the Lor platform does

Mirendil’s first product is Lor, which automates many of the steps in a model training run. The company says the platform spreads that work across hundreds of AI agents.

Building a large language model involves a lot of trial and error. Researchers make several slightly different versions of an algorithm and test them against each other to find the one that performs best. Mirendil says Lor can run that process on its own and repair errors that come up during testing.

A model built by one researcher

To test the approach, Mirendil asked a single researcher earlier this year to build a coding model meant to compete with leading open-source language models. Such projects normally need a team of specialists. According to the company, the researcher finished without help from colleagues, in about 3.5 months and at a cost of $10 million.

Mirendil says its agents wrote more than 99% of the code behind that model, called Mirendil-Base. The agents set up the infrastructure, picked a training algorithm and gathered training data from the web, then adjusted the dataset and the model’s design to improve the quality of its output.

Mirendil-Base is the foundation for Fern, a coding model the company unveiled last month to handle programming problems that AI labs run into often. Mirendil says Fern can take in 10 million input tokens in a single prompt, 10 times more than Claude Opus 5.5, or roughly a half-million lines of code. Tokens are the small chunks of text a model reads.

Fern runs on a version of Mirendil-Base with 100 billion to 500 billion parameters, the internal settings a model learns during training. Late last month the company previewed two upgraded versions of Mirendil-Base with between 500 billion and 10 trillion parameters.

What it means for investors and the AI market

With Google in the round, a company whose own DeepMind lab builds AI models now holds a stake in a startup selling automation to AI labs. Mirendil’s case to customers rests on cost and speed: it says a model that would normally take a team was built by one person for $10 million.

What comes next

Mirendil says it plans to use the larger versions of Mirendil-Base to develop more capable language models. It has not said when those models will be released.

Source: SiliconANGLE