Live·Open questions in longevity research
All news
Scientific Computing

Mirendil raises a billion dollars for tools to help small teams develop AI

16 November 2026· 261116006

Mirendil raises a billion dollars for tools to help small teams develop AI

The company announced the funding on November 12. Investment firms Kleiner Perkins and Clearlake Capital led the round, with Google among the participants. Mirendil is developing its own models and software to write research code, run computational experiments, and analyze the results.

Training a new neural network requires preparing data, choosing a model architecture, and coordinating computations across many processors. These tasks are usually divided among specialists. Mirendil tested whether one experienced researcher could train a model with the help of AI agents, programs that take a task, execute commands on their own, and use the results to decide what to do next.

In its account of the experiment published on October 21, the company reports that one experienced researcher working with agents trained Mirendil-Base, a model for working with code, in three and a half months. The researcher and agents designed the model architecture and developed the data preparation and training software themselves, using existing general-purpose software tools. According to Mirendil, the agents wrote more than 99% of the code, ran experiments, and investigated failures; the researcher decided what to try next. Compute costs totaled less than 10 million dollars, including preliminary experiments, the final training run, data processing, and running the agents themselves.

Mirendil offers other researchers Lor, an environment for managing agents and computing resources, currently available by invitation. The company gives an example task: reproduce the calculations from a scientific paper on eight graphics processing units and explain any discrepancies with the numbers in its table. According to Mirendil, an agent can divide the task among dozens of other agents, request computers for the calculations, and run programs; Lor routes the commands to the selected computers. Hundreds of agents can work in parallel for several days, using models from Mirendil and other developers.

“People decide what to try and what to trust.”

Mirendil cofounder Harsh Mehta built the first version of the research automation system at the AI company Anthropic, initially working alone. Mirendil now wants to make these tools available to outside laboratories so they can train their own AI on their own data for their own tasks. The company envisions small teams taking on more tasks by adding agents and computers to run the calculations.

Mirendil intends to use its own models to write research code and analyze experiments as it develops the next generation of AI. The company describes the expected effect as follows:

“Each model that gets better at building AI makes the next model cheaper and faster to build.”

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-agents#model-training#research-automation#scientific-computing#mirendil#series-a