Mirendil’s $1B Series A puts automation at the heart of AI model development

Mirendil Inc. has raised $1 billion in a Series A round led jointly by Kleiner Perkins and Clearlake Capital, with Google and half a dozen other investors joining. The startup, founded last year by two former Google DeepMind and Anthropic researchers, is betting that automation can replace the trial-and-error grind of building large language models. Its flagship product, Lor, doesn’t just optimize training runs—it turns the process into a self-correcting workflow where AI agents handle everything from infrastructure setup to architecture refinement.

The funding comes as the AI development pipeline remains one of the last frontiers where labor and iteration still dominate. Most LLMs require teams of specialists to tweak algorithms, curate datasets, and debug errors. Mirendil claims to have broken that mold. In a test earlier this year, the company tasked a single researcher with building a coding-focused LLM competitive with leading open-source models. Using Lor, the researcher completed the project in 3.5 months for $10 million—without relying on colleagues. The platform’s agents wrote 99% of the code, selected training algorithms, and even refined the dataset to improve output quality.

The result was Mirendil-Base, a 100-billion- to 500-billion-parameter foundation model. That model underpins Fern, a specialized coding assistant unveiled last month. Fern can process 10 million input tokens per prompt—ten times more than Claude Opus 5.5—effectively handling half a million lines of code at once. The company has already previewed upgraded versions of Mirendil-Base, scaling from 500 billion to 10 trillion parameters, suggesting it’s positioning itself as both a developer tool and a foundational infrastructure player.

Startups like Zenithon (covered here on 30 September) and Dream Labs (4 November) are targeting niche applications, but Mirendil’s approach focuses on the bottleneck of *how* models are built. Its CEO, Behnam Neyshabur, is a co-inventor of SAM, a widely used algorithm for zero-shot generalization, while CTO Harsh Mehta’s background in pre-training at DeepMind and Anthropic lends credibility to its claim of addressing the “trial-and-error” phase head-on.

What sets Mirendil apart isn’t just the scale of its funding but the scale of its ambition. Most AI startups either build models or optimize hardware; Mirendil is aiming to redefine the *process* of model creation itself. The company’s demo of a solo researcher outpacing a team on a $10 million project—while still a proof of concept—hints at a future where AI development resembles software engineering more than academic experimentation.

The open question is whether Lor’s automation can scale beyond coding-focused models. Mirendil hasn’t disclosed broader use cases, but the preview of trillion-parameter variants suggests it’s eyeing enterprise or scientific applications where iterative refinement is critical. If successful, it could redefine not just how models are trained but who gets to train them—and at what cost.

For founders and operators watching this space, the takeaway is clear: the next wave of AI infrastructure isn’t just about compute or data. It’s about turning the messy, manual work of model development into something that can be delegated to agents. The $1 billion round isn’t just capital—it’s a vote of confidence in that vision.

Sources: siliconangle.com