Mirendil raises $1B in funding to speed up AI model development

Mirendil Inc., a startup that makes software for accelerating artificial intelligence development, has raised $1 billion in early-stage funding. The company said Thursday that Kleiner Perkins and Clearlake Capital jointly led the Series A round. More than a half-dozen other investors participated, including Google LLC.

Mirendil was founded last year by two former researchers from Google DeepMind and Anthropic PBC. CEO Behnam Neyshabur is a co-inventor of SAM, an algorithm that improves how well AI models handle tasks they were not specifically trained for. CTO Harsh Mehta worked on pretraining projects during his time at Google DeepMind and Anthropic.

The company's first product, Lor, automates many tasks involved in model training runs by distributing work across hundreds of AI agents, according to Mirendil. Large language model development typically requires extensive trial and error, with researchers producing many slightly different versions of an algorithm and comparing them to find the best performer. Mirendil says Lor can automate that process and automatically correct errors that arise during evaluation.

Earlier this year, Mirendil asked one of its researchers to build a coding model capable of competing with leading open-source LLMs, a task usually handled by teams of specialists. According to the company, its automation technology let the researcher complete the project alone. The effort cost $10 million and took about 3.5 months.

Mirendil says its agents wrote more than 99% of the code behind the resulting model, Mirendil-Base. The agents set up infrastructure, selected a training algorithm and gathered training data from the web. They then tuned both the dataset and the model's architecture to improve output quality.

Mirendil-Base is the foundation of Fern, a specialized coding model the company unveiled last month. According to Mirendil, Fern is designed to automate programming challenges commonly faced by AI labs. It can process 10 million input tokens per prompt, or 10 times more than Claude Opus 5.5, roughly equivalent to about a half-million lines of code.

Fern is based on a version of Mirendil-Base with 100 billion to 500 billion parameters. Late last month, the company previewed two upgraded versions of that model with between 500 billion and 10 trillion parameters. Mirendil plans to use those algorithms to develop more capable LLMs.

More company and startup news from TechManNews.