Brands
Latest top stories
Start-ups
Technology

Mirendil raises $1 billion to automate AI research

16 November 2026

 

Mirendil, a US startup developing tools to automate AI research, announced its $1 billion Series A on 12 November. Kleiner Perkins, which co-led the $200 million seed round with Andreessen Horowitz, returned to lead the Series A alongside Clearlake Capital. Google, Battery Ventures and Pear VC were among the other participants. The Wall Street Journal, citing people familiar with the deal, reported that the round valued the company at $5 billion. The two announced rounds account for the $1.2 billion figure. Mirendil also disclosed a subsequent investment from NVIDIA when it introduced the company in June, without giving an amount.

 

The team behind the funding

 

Kleiner Perkins’ account of its seed investment traced the idea to the founders’ work at larger labs. Chief executive Behnam Neyshabur was a senior staff research scientist at Google DeepMind, where he co-led Blueshift, the team that developed the maths-focused Minerva model. He and co-founder Harsh Mehta worked on mathematical reasoning and coding in Gemini before moving to Anthropic. There, Neyshabur co-led the Discovery team, whose goal was to build an AI scientist, while Mehta started and led an effort to automate AI research and development.

Shayan Salehian brought experience from xAI, where he worked on Grok and the infrastructure that runs AI agents. Tara Rezaei, a Massachusetts Institute of Technology graduate, had been a student researcher at OpenAI. Their combined experience helps explain Kleiner Perkins’ interest. It saw a team that could build models and use them to speed up the experiments that inform the next generation. Mirendil’s October report showed how much of that work its agents could already take on under a researcher’s direction.

 

How AI agents helped train Mirendil-Base

 

During pretraining, a model learns patterns from large quantities of data. Behind a training run are decisions about which data to use, how to organise the work and what to change when an experiment stalls. In Mirendil’s project, the researcher chose which ideas to pursue and which results to trust, while the agents ran experiments and diagnosed failures. Mirendil says they wrote more than 99 per cent of the project’s code. The result was Mirendil-Base, a model aimed at the coding tasks AI researchers encounter.

Mirendil says the project took three and a half months and used less than $10 million in computing resources. That figure covers the agents’ usage, data processing, preliminary experiments and the final pretraining run. In a company-run test of five models, Mirendil-Base averaged the second-best score at predicting code across seven external repositories. The result gives a point of comparison for the base model on that task. A benchmark of whether it could deliver working software fixes would require a different evaluation.

 

Fern reads the logs and Lor runs the experiments

 

Further training turned Mirendil-Base into Fern, a model designed for coding and research. Mirendil says Fern can accept up to 10 million tokens, the chunks into which it divides text and code. For a researcher investigating a failed run, that could mean giving the model several days of logs alongside the code and results. The size of the context window measures how much the model can take in. How consistently it can pick out the relevant evidence across that volume is a separate test.

Lor is Mirendil’s platform for running AI agents with models such as Fern. Mirendil says it uses the platform internally and that it can coordinate hundreds of agents working in parallel for days. It also supports models from other developers. A researcher can follow the individual agents’ progress and stop them as needed, then examine the results before deciding which line of enquiry to pursue. Lor is free during its invite-only beta.

 

The economics of a smaller AI lab

 

Mirendil wants small teams in fields such as materials science and robotics to use Lor to develop their own AI models. A robotics team, for example, could set out to teach a model to recognise when a machine is starting to drift out of alignment. The engineers would decide what counts as an early warning and how to test it on the machine, while Lor’s agents could write the training code and run experiments on the team’s data.

Delegating the engineering still leaves a compute bill. Mirendil has a multi-year agreement to use Google Cloud’s own AI accelerators and NVIDIA infrastructure, while Baseten says it supports some of the company’s reinforcement learning workloads. A small team considering Lor will need to know how much computing power a project would use and what it would cost once the free beta ends.

End of the day, investors have backed a team that knows how to train large AI models and has shown how much of the engineering it can delegate. The wider promise is that a small team could develop a model around a problem it knows firsthand. For the robotics engineers, that would mean getting an early warning they can trust on the factory floor, at a cost that lets them refine the model as the machines change.

 

 

 

Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.

👉 https://buymeacoffee.com/movetheneedle.news