Mirendil Inc., a provider of software that speeds up artificial intelligence development, has raised $1 billion in early-stage funding.
The startup stated on Thursday that Kleiner Perkins and Clearlake Capital jointly led the Series A investment. They were joined by more than a half-dozen others including Google LLC.
Mirendil was launched last year by two former Google DeepMind and Anthropic PBC researchers. Chief Executive Officer Behnam Neyshabur is a co-inventor of SAM, a popular algorithm that makes AI models better at tasks they were not specifically trained to perform. Chief Technology Officer Harsh Mehta, in turn, spent his time at Google DeepMind and Anthropic working on pretraining projects. Pretraining is a term for the initial phase of the AI development workflow.
Mirendil’s first product, Lor, is a platform that automates many of the tasks involved in model training runs. The company says that the software does so by splitting work among hundreds of AI agents.
Large language model development involves a significant amount of trial and error. Researchers create multiple versions of an algorithm that differ from one another in small ways and compare them to find the most performant variant. According to Mirendil, Lor can automate that workflow. If an error crops up during the evaluation, the platform fixes it automatically.
Earlier this year, Mirendil asked one of its researchers to build a coding model that can compete with leading open-source LLMs. Usually, such projects are carried out by teams of specialists. Mirendil says that its automation technology enabled the researcher to build the model without any help from colleagues. The project cost $10 million and took about 3.5 months to complete.
The company’s agents wrote more than 99% of the code that powers the model, which is called Mirendil-Base. They set up infrastructure, found a suitable training algorithm and collected training data from the web. The agents then refined both the training dataset and Mirendil-Base’s architecture to boost output quality.
The LLM is the foundation of a specialized coding model called Fern that the company unveiled last month. According to Mirendil, it’s designed to automate programming challenges commonly encountered by AI labs. Fern can process 10 million input tokens per prompt, or 10 times more than Claude Opus 5.5. That’s the equivalent of 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 the latter model with between 500 billion and 10 trillion parameters. Mirendil plans to use those algorithms to develop more capable LLMs.
Image:Unsplash