Mirendil makes a simple promise to people who build AI models: hand the grunt work to a crowd of agents. Its product, Lor, runs hundreds of AI agents at once that write code, run experiments and keep iterating for days. On 12 November the lab said it had raised $1 billion in a Series A led by Kleiner Perkins and Clearlake Capital. It did not say what the company is now worth.

What Lor actually does

You give Lor’s agents a research chore and they carry it out on real machines. Mirendil’s examples read like an AI lab’s to-do list: “Run these 12 ablations overnight and write a report on the results.” An ablation is a test where you remove one part of a model to see what it was doing.

The agents bring their own computing power: they request servers, run the job and hand them back, with no cloud account needed. Lor runs models such as Claude Fable and GPT-6 Astra, plus open models like Kimi and DeepSeek. We looked at what Anthropic’s Claude Fable 5.2 can build and what it costs earlier today.

Mirendil says people stay in charge: agents only work on machines you have added, and you can stop any of them at any time. Lor is in private beta with a waitlist.

The claim behind the round

Mirendil’s biggest claim is about its own work. In October it wrote that one researcher working with agents pretrained a coding model, Mirendil-Base, “in 3.5 months, with less than $10M in compute.” Pretraining is the first and costliest stage of building a model. The company says agents wrote over 99% of the code.

Those are the company’s numbers, and nobody outside has checked them. Tech Times points out that no independent evaluation of any Mirendil model has been published. The under $10 million bill also covers only the smallest model in the family. On Mirendil’s own code prediction test, Mirendil-Base came second, behind Kimi-K3.

Who is paying, and how much

Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 joined the round, according to Mirendil. Google wears two hats: it is also a supplier, after Mirendil announced a multi-year deal to train on Google Cloud in August.

The people are a big part of the bet. SiliconANGLE reports that chief executive Behnam Neyshabur and chief technology officer Harsh Mehta both came from Google DeepMind and Anthropic. In June, Mirendil launched with a $200 million seed round led by Andreessen Horowitz and Kleiner Perkins.

In September, Bloomberg reported that Mirendil was in talks to raise up to $1 billion at a $5 billion valuation. Those were talks, not confirmed terms, and Mirendil’s announcement gives no valuation. The $1 billion round is twice the size of Genesis AI’s $500 million Series A for a glove-trained robot.

Agents that start their own servers raise a question about guardrails. Tech Times notes that the Lor page describes no sandboxing or spending limits. The case where California sued OpenAI over test agents that hit Hugging Face has nothing to do with Mirendil, but it shows what can go wrong.

What to do this week

If your team trains models and the experiment queue is the bottleneck, join the Lor waitlist. Before any pilot, ask what it will cost after the beta, what cap stops agents from starting extra servers, and which of your machines they can reach. Then test it on a job you already ran by hand and compare.

Sources

Martin Scott

Martin Scott is the editor for business and policy at AI Observer, covering startups and deals, regulation, health and science, and AI agency profiles. He follows the money and the legal text, and separates what a rule says from what people claim it says. His style is analytical and plain-spoken.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.