Mirendil Inc. Secures $1 Billion in Early-Stage Funding to Accelerate AI Development
Mirendil Inc., a software provider focused on expediting artificial intelligence development, has successfully raised $1 billion in early-stage funding. The announcement was made on Thursday, highlighting the company’s commitment to democratizing advancements in frontier AI research and development.
The funding round was led by Kleiner Perkins and Clearlake Capital, supported by additional investments from over a half-dozen other participants, including Google LLC. This financial backing positions Mirendil to enhance its offerings and broaden its impact within the AI landscape.
Founded last year by researchers from Google DeepMind and Anthropic PBC, Mirendil is led by Chief Executive Officer Behnam Neyshabur, who is a co-inventor of the widely recognized algorithm known as SAM. This algorithm enhances the performance of AI models in areas outside of their original training. Chief Technology Officer Harsh Mehta has extensive experience in pretraining projects and provides critical expertise to the startup’s innovative direction.
Mirendil’s flagship product, Lor, is designed to automate many processes involved in model training. By utilizing hundreds of AI agents, the software streamlines tasks typically requiring significant manual effort, thereby accelerating the overall development timeline.
Development in the realm of large language models entails numerous iterations, as researchers create varying versions of algorithms and assess their relative performance. Mirendil’s Lor platform can automate this iterative workflow, with capabilities to automatically address errors that arise during evaluations.
Earlier this year, Mirendil tasked one of its researchers with developing a coding model that could rival leading open-source large language models (LLMs). Traditionally, such projects rely on expert teams; however, Mirendil reported that its automation technology enabled the researcher to execute the initiative independently. The project, which required $10 million in funding and spanned about 3.5 months, exemplifies the potential efficiency gains through Mirendil’s approach.
Significantly, over 99% of the code for the newly developed model, named Mirendil-Base, was authored by AI agents. They were responsible for establishing the necessary infrastructure, selecting appropriate training algorithms, and gathering training data from online sources. These agents also played a crucial role in refining the dataset and optimizing the architecture of Mirendil-Base to enhance output quality.
The Mirendil-Base model serves as the foundation for Fern, a specialized coding model that the company introduced last month. Fern is engineered to tackle common programming challenges faced by AI laboratories and boasts the capability to process 10 million input tokens per prompt, surpassing Claude Opus 5.5 by a factor of ten. This capacity equates to handling approximately half a million lines of code.
Fern operates on a differentiated version of Mirendil-Base, which contains between 100 billion and 500 billion parameters. In a recent preview, Mirendil revealed plans for upgraded versions of the model, featuring between 500 billion and 10 trillion parameters. These advancements signal the company’s ongoing commitment to developing increasingly powerful large language models for the AI community.

