aipopdb

19 Nov 2026  ·  9,164 works · 1,132 news indexed

News / Mirendil raises $1 billion Series A to automate AI research

Mirendil raises $1 billion Series A to automate AI research

AIPopDB Newsroom · via siliconangle.com · · 2 min read

Read the original ↗
Image for Mirendil raises $1 billion Series A to automate AI research

Image: Unite.AI

Watch on YouTube ↗

Mirendil, a startup that builds software and models to speed up AI research and development, said on 12 November 2026 that it has raised a $1 billion Series A. Kleiner Perkins and Clearlake Capital led the round jointly. Google, Battery Ventures, Doerr Capital, Gil & Co, M13, Pear VC and S32 joined, along with other new and existing backers. The announcement gives no valuation.

“We believe making AI R&D more accessible is the most direct path to AI that solves humanity's most pressing problems across science and technology.”

Mirendil, in its funding announcement · ultrathink.ai ↗

The company is run by CEO Behnam Neyshabur, who co-invented SAM, an algorithm that helps models handle tasks they were not trained for, and CTO Harsh Mehta, who worked on pretraining at Anthropic. Mirendil's stated aim is to let one person do what took a team, and a team do what took an institution.

The first product, Lor, splits model-training work among hundreds of AI agents and fixes errors found during evaluation. It was introduced on 28 October as an invite-only private beta, free for now, on macOS, Linux and the web, and Mirendil says it runs on Lor daily. As proof, it says one researcher built the coding model Mirendil-Base in about 3.5 months for under $10 million all-in, with agents writing over 99% of the code, setting up infrastructure and gathering web training data. It says the model used only public code, web pages and datasets, with no distillation or synthetic data.

Mirendil-Base sits at 100 billion to 500 billion parameters, and larger versions of 500 billion to 10 trillion are in training. Fern, post-trained from it, takes up to 10 million input tokens, about half a million lines of code, which SiliconANGLE puts at ten times Claude Opus 5.5. Fern 0.1 access is listed as coming soon, as are versions of Lor for Windows, iOS and Android.

Mirendil
AI company
United States

About Mirendil

Mirendil is a frontier AI lab that says it builds systems that excel at AI research and development, training models for the job and redesigning the lab around them.

The company announced itself on 24 June 2026 with a $200 million seed round at a $1 billion valuation, led by Andreessen Horowitz, with Kleiner Perkins and Nvidia among the investors. On 6 August 2026 it announced a multi-year Google Cloud partnership for training and inference on TPUs and Nvidia hardware. On 21 October 2026 it reported that one researcher with AI agents pretrained Mirendil-Base in 3.5 months for under $10 million, and a week later it introduced the Lor agent platform.

2

Models indexed

Models on AIPopDB

Fern
Mirendil-Base
See Mirendil on AIPopDB

Unite.AI calls the central test whether outsiders can repeat the one-researcher result. Mirendil plans to use the larger Mirendil-Base models to build more capable language models.