Mirendil raises $1B to expand agent-led AI research

• Mirendil announced a $1 billion Series A led by Kleiner Perkins and Clearlake Capital, five times its disclosed seed round.
• The company says one researcher and AI agents pretrained a coding model in 3.5 months using less than $10 million in compute. That is not a comparison of total project costs.
• Lor, its research-agent platform, supplies compute and runs outside models, but remains in a free private beta with commercial pricing undisclosed.

Mirendil, a company building models and software to help scientists and engineers develop AI, said on November 12 that it had raised $1 billion. The financing backs its bid to let smaller teams build their own models by handing much of the engineering to agents, while people direct the research and judge the results.

The company-issued release describes a model-building experiment, a long-context model and a platform for running hundreds of agents. The distinction that matters for prospective users is between needing fewer specialists and being able to afford the resulting research. Mirendil’s experiment still required substantial compute and an experienced researcher; its product has yet to establish commercial costs for outside laboratories.

Mirendil disclosed a $200M seed round and a $1B Series A. These amounts exclude its separately announced, unquantified NVIDIA investment. Source: Mirendil founding announcement.

A billion-dollar round follows a $200 million seed

Mirendil formally introduced itself on June 24 with a founding team of 20 researchers and engineers from AI developers Anthropic, xAI, Google DeepMind and OpenAI. Its founding announcement disclosed a $200 million seed round led by venture investors Andreessen Horowitz and Kleiner Perkins, followed by an investment from NVIDIA and others. The Series A alone is five times that disclosed seed amount; the subsequent NVIDIA investment was not quantified there.

The team’s background is relevant to the task it is trying to automate. In its June investment rationale, seed backer Andreessen Horowitz cited team members’ experience working on AI for science at Google and Anthropic and building the first version of Anthropic’s research automation platform.

The investor also said most major AI labs already had versions of such a platform for internal use. Its case for an independent company rests partly on its argument that those labs have little economic incentive to release their platforms to competitors. That frames Mirendil’s opportunity as opening access to research tools, rather than inventing research automation.

Kleiner Perkins is returning as a lead. Clearlake Capital, a global investment firm spanning private equity and credit, is a new investor, according to a transaction report that also says Mirendil issued convertible preferred stock. Mirendil’s announcement does not disclose a Series A valuation, ownership percentages or board arrangements.

The company’s funding announcement names Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 among the other participants.

Google is also an infrastructure supplier. In August, Mirendil announced a multiyear Google Cloud partnership for training, inference and research workloads using Google’s tensor processing units and NVIDIA computing infrastructure. Mirendil says the arrangement includes training clusters co-designed across computing, storage and networking. The funding therefore deepens a relationship that already helps supply the machines its agents need; the announcements provide no contract value or unit compute prices.

Fewer engineers does not mean autonomous research

In its October 21 research disclosure, Mirendil said one researcher working with agents pretrained Mirendil-Base, a coding-specialized model, from scratch in 3.5 months using less than $10 million in compute. The figure includes research experiments, the final training run, agents’ token consumption and data processing. It is not a disclosed total business cost or a measured saving against an equivalent conventionally staffed project.

Mirendil says agents wrote more than 99% of the code, ran experiments and diagnosed failures. The researcher chose what to try, judged what to trust and intervened. They built training infrastructure, model architecture, the training algorithm and data pipelines on standard libraries including PyTorch and JAX, using open-source projects as references. Training used public code, web pages and datasets, without purchased additional data, distillation or synthetic data.

That is a claim about reducing the engineering labor needed for pretraining, not a demonstration of a model building itself. Mirendil acknowledges that AI agents already do a growing share of model-development work inside other frontier labs. It identifies compute, data and human judgment as remaining constraints, and describes its own experiment as the work of an experienced researcher.

The alternative for customers also matters. Mirendil says most organizations currently rent general-purpose models and adapt them. Its ambition is to make building a base model accessible to teams such as robotics laboratories or hospitals. Its experiment does not establish when that approach becomes cheaper or more useful than adapting an existing model.

Company-reported code-prediction averages across seven external repositories, from Mirendil. Lower is better; Kimi-K3 and GLM-5.3 are post-trained models. Source: Mirendil pretraining research.

The coding benchmark has a stronger rival

Mirendil’s comparison of coding models measures code prediction, using cross-entropy in bits per byte: lower scores mean better prediction. This permits comparison across models with different tokenizers, but does not directly measure completed research projects, engineering reliability or customer costs.

On seven external code repositories published after the models were trained, the company reported these average scores:

In that company-run comparison, Kimi-K3 scored better than Mirendil-Base on six repositories and tied on the seventh. Mirendil also notes that Kimi-K3 and GLM-5.3 are post-trained models, which generally score slightly worse on this metric than their base versions. The evidence supports a competitive code-prediction result, with a stronger rival in the same test, rather than a general claim of research leadership.

Scale is another qualification. Mirendil places its completed Base model in a broad range of 100 billion to 500 billion parameters. Its October research post described Base-Large, at 500 billion to 2 trillion parameters, and Base-XL, at 2 trillion to 10 trillion, as still in pretraining. The financing announcement supplies neither completed results nor final costs for those larger models.

Lor packages agents, models and compute

Mirendil says it subsequently applied post-training to Base to create Fern, a model that can hold up to 10 million tokens of context for work involving long agent histories, training logs and codebases. A report describing Mirendil’s October 27 product update said Fern access was forthcoming at that point. Lor’s later announcement lists Fern among its supported models; that does not establish general availability beyond the platform’s private beta.

Lor, introduced on October 28, is software for coordinating agents that write code and run experiments for days or weeks. In its product announcement, Mirendil says Lor supplies its own compute: agents request nodes, configure them, run work and release them, without users needing a separate cloud account. It also runs outside providers’ models, so customers need not rely exclusively on Mirendil’s model quality.

The company describes agents running in the cloud and sending commands to a chosen machine, which can be a Lor-managed computer or a user’s existing machine. Users can stop agents and remove machines; agents operate only on machines the user has added. This separates the place an agent runs from the hardware on which it performs work.

In one company-reported demonstration, 102 agents built environments from 100 PyTorch pull requests, averaging eight minutes per agent. That is a specific engineering workload, not the elapsed time for an entire research program. Lor is available on macOS, Linux and the web, with access restricted to invitations through a waitlist.

Mirendil’s Lor interface during a company-reported demonstration using 102 agents to build environments from 100 PyTorch pull requests. Source: Mirendil Lor announcement.

What outside laboratories still need to know

Mirendil says it runs on Lor daily, and use is free during the private beta. Bundling compute removes one setup requirement, but a free trial supplies no basis for calculating what sustained agent activity and experiments will cost under commercial terms.

The next consequential evidence is outside use: how much expert supervision successful projects require, what their full costs are, and when building a model offers an advantage over adapting one already available. Completed results for the larger models would test a separate scaling claim. The financing gives Mirendil resources to pursue those tests; pricing and reproducible customer results will determine how far its promised access extends.