1. Home
2. News
3. Business

Mirendil raises $1 billion, saying one researcher with AI agents pretrained its first model

Its case is a model it says one researcher built with agents writing over 99% of the code.

By Think Facility · November 13, 2026 · AI business AI research Save

Mirendil, a startup that wants AI to do most of the work of building AI, said on November 12, 2026, that it has raised $1 billion. Its announcement says Kleiner Perkins and Clearlake Capital led the Series A, with Google among the other investors.

The company only came out of stealth in June. Back then it described itself as 20 researchers and engineers from places including Anthropic, xAI, Google DeepMind and OpenAI, with a $200M seed round behind it.

Raised

$1B in a Series A

Led by

Kleiner Perkins and Clearlake Capital

Seed round

$200M, led by Andreessen Horowitz and Kleiner Perkins

First model

Mirendil-Base, pretrained by one researcher in 3.5 months

Compute for it

less than $10M, all-in

The claim behind the money

Mirendil's pitch rests on one experiment. In an October 21 post, it said a single researcher, working with AI agents, pretrained a competitive coding model from an empty repository. That's the kind of job that has taken dozens of specialists at open-model efforts and hundreds at frontier labs, by the company's own account.

The model, Mirendil-Base, sits somewhere between 100 billion and 500 billion parameters. The company says it trained only on public data, with no distillation and no bought or licensed data. Two bigger versions are still training: Large was about 80% of the way through on October 21, and XL about 19%.

These are Mirendil's own numbers, and nobody outside the company has checked them in anything we saved. Its comparison chart measures how well each model predicts code from Mirendil's own repositories and seven outside ones. On the outside set, Kimi-K3 averaged lower (better) than Mirendil-Base. Mirendil-Base beat the other three base and post-trained models it was shown against.

How close is self-improving AI?

The company hedges this part. It says Mirendil-Base didn't build itself; a person did, with AI tools. But it also says its next model will be built partly by Mirendil's own models.

Mirendil cites published internal measurements from Anthropic and OpenAI showing agents doing a growing share of AI-building work. Its product, Lor, is meant to sell that way of working to everyone else. It runs hundreds of agents at once for days, on frontier models such as Claude Opus and GPT-6 Sol as well as Mirendil's own Fern, which holds up to 10M tokens of context. Lor is in private beta and free for now.

What it would mean for jobs

The company's argument about work is blunt. Organizations have grown by hiring, it says, and compute now gives them another way: a team can take on more by adding compute. Its Series A post says it wants a world where one person can take on what used to need a team.

Mirendil names who it thinks gains. A robotics lab or a hospital, for instance, could train a model on its own data instead of renting a general one. It doesn't say what happens to the specialists whose work the agents take over, and none of its posts gives a number for jobs.

The company says Mirendil-Base-XL, which it calls a frontier-class model, is still in training, and it'll share updates on it.