Mirendil Raises $1B Series A to Automate AI Training With Fleets of Agents

Mirendil Inc. has raised $1 billion in a Series A funding round to develop software that automates the engineering workflows behind artificial intelligence model training.

The investment was jointly led by Kleiner Perkins and Clearlake Capital, with participation from Google, Battery Ventures, Doerr Capital, Gil & Co, M13, Pear VC, and S32, according to a company announcement on Thursday covered by SiliconANGLE (https://siliconangle.com/2026/11/13/mirendil-raises-1b-in-funding-to-speed-up-ai-model-development/).

Founded in 2025 by former Google DeepMind and Anthropic researchers, the San Francisco startup aims to reduce the engineering headcount required to train frontier models. Frontier pretraining runs routinely demand large teams of specialists to provision infrastructure, monitor loss curves, assemble datasets, and intervene when runs fail. Mirendil aims to substitute much of that manual intervention with coordinated software agents.

The company's core product, Lor, is an orchestration platform currently in private beta that splits model training workloads across hundreds of AI agents. In large language model research, engineers systematically test subtle variants of algorithms and architectures to discover optimal configurations. Mirendil claims Lor automates this search process, evaluating iterations and self-correcting training errors when runs stall.

As an internal test of the platform, Mirendil tasked a single researcher with training an open-source-competitive coding model from scratch, called Mirendil-Base. According to the company, autonomous agents wrote more than 99% of the code, provisioned compute clusters, collected web training datasets, and tuned the architecture. Mirendil reported that the model took three and a half months to complete and consumed under $10 million in compute.

Mirendil subsequently post-trained Mirendil-Base into Fern, a domain-specific coding model introduced last month to tackle programming problems faced by AI labs. Mirendil claims Fern can handle an input context window of 10 million tokens—roughly 500,000 lines of code. Fern is built on a version of Mirendil-Base sized between 100 billion and 500 billion parameters; the startup has also previewed model checkpoints scaling from 500 billion up to 10 trillion parameters.

Mirendil Chief Executive Behnam Neyshabur previously co-invented Sharpness-Aware Minimization, an optimization procedure designed to improve model generalization by seeking flatter loss minima. Chief Technology Officer Harsh Mehta previously led pretraining initiatives at Google DeepMind and Anthropic.

The startup intends to use the new financing to widen access to Lor and fund the substantial compute footprints required to train its planned multi-trillion-parameter systems.