Mirendil Raises $1 Billion Series A to Put a Frontier AI Lab's Workload in the Hands of One Researcher
Kleiner Perkins and Clearlake Capital co-led the round for the year-old startup, whose Lor platform runs hundreds of AI agents for days at a time to write code, run experiments and provision their own GPUs.
Published
Last updated / Fact-checked on
Illustrative photo of a data center; it does not show Mirendil's facilities or compute.
Mirendil, a San Francisco startup that wants AI agents to do the hands-on engineering behind frontier AI research, has raised $1 billion in a Series A round. Kleiner Perkins and Clearlake Capital co-led the financing, the company said in a post published at 9 a.m. Pacific on Thursday, November 12, with participation from Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32, among other new and existing investors.
A billion dollars is an unusually large first priced round, even by the standards of a year in which AI startups have raised at record pace. Mirendil did not disclose a valuation. The Tech Startups funding roundup noted that earlier reports of talks at a possible $5 billion valuation should not be treated as the final terms. Law firm Sidley Austin said it represented Clearlake on its co-lead investment.
Who is behind it
SiliconANGLE reported that Mirendil was founded last year by two researchers who came from Google DeepMind and Anthropic. Chief executive Behnam Neyshabur is a co-inventor of SAM, a widely used training technique that helps models generalize to tasks they were not specifically trained on. Chief technology officer Harsh Mehta worked on pretraining, the first and most compute-hungry phase of building a large language model, at both labs.
The company describes its mission in sweeping terms. "We believe making AI R&D more accessible is the most direct path to AI that solves humanity's most pressing problems across science and technology," Mirendil wrote in its announcement. Its pitch is that once agents handle the tedious parts of research, team size stops being the limit on what can be built: "one person can take on what used to need a team, a team what used to need an institution."
The proof point: a model built by one researcher
To back that claim, Mirendil points to its own first model. According to the company, a single researcher working with agents pretrained Mirendil-Base, a coding model, from scratch in three and a half months using less than $10 million in compute. SiliconANGLE, relaying the company's description, reported that the agents wrote more than 99% of the code behind the model, set up the infrastructure, chose a training algorithm, collected training data from the web, and then refined both the dataset and the model's architecture.
Mirendil then post-trained that base into Fern, a model that can hold up to 10 million tokens of context. The company says that is enough to keep days of an agent's work in view at once. SiliconANGLE put the figure at ten times the context of Anthropic's Claude Opus 5.5, or roughly half a million lines of code, and reported that Fern is built on a version of Mirendil-Base with between 100 billion and 500 billion parameters. Late last month, the outlet said, Mirendil also previewed two larger versions of the base model in the 500 billion to 10 trillion parameter range, which it plans to use to build more capable systems.
Those numbers have not been independently benchmarked in the sources DailyMech reviewed. Mirendil's announcement calls Mirendil-Base "competitive," and SiliconANGLE described the goal as a model that can compete with leading open-source LLMs, but neither offered third-party evaluation results.
What Lor actually does
The product the money is meant to scale is Lor, which Mirendil introduced on October 28 and which remains in private beta by invitation. The company describes it as a system "to do the work of an entire AI lab."
In Lor, agents run in the cloud but send their commands to whichever machine the user chooses: a Lor-managed cloud computer, a laptop, or any other always-on machine the user has added. That separation is what lets work continue after a user closes the laptop. Agents can request compute nodes, set them up, run training jobs or evaluation sweeps, and release the hardware when they finish, without the user managing a cloud account.
The other core idea is parallelism. One agent can split a job across dozens of others, and a user can run hundreds across projects while watching progress and stopping any agent at any time. Mirendil gives two examples: in one run, 102 agents built training environments from 100 PyTorch pull requests, averaging eight minutes each; in another overnight run, agents found and fixed security bugs in the LiteLLM project by morning. The company also lists tasks a user might hand off, such as running a dozen ablation experiments overnight and writing up the results, or reproducing a table from a research paper on eight GPUs and explaining where the numbers differ.
Lor is not tied to Mirendil's own models. The product page lists frontier models from Anthropic, OpenAI, Google, xAI and Meta alongside open-weight families from Z.ai, Moonshot AI and DeepSeek, with Fern as the in-house option. It runs on macOS, Linux and the web, and it is free during the beta. Windows, iOS and Android versions are listed as coming soon. Mirendil says it uses Lor internally every day.
Why it matters
Mirendil is betting on a specific idea: that the scarce resource in AI research is no longer just GPUs but the skilled engineers who run, debug and iterate on experiments, and that agents can stand in for much of that labor. If that holds, the advantage of the largest labs, which can afford hundreds of research engineers, could narrow. That is the argument behind the "democratize frontier AI R&D" language in the announcement.
The round also fits a broader shift in where agent money is going. Investors have moved from chat assistants toward systems that run long, unattended jobs with real resources. Lor's ability to spin up its own compute and act on user-supplied machines raises the stakes on control. Mirendil says agents operate only on machines a user has added, which can be removed at any time, and that people still decide "what to try, and what to trust."
Key questions remain open. Mirendil has not published independent evaluations of Mirendil-Base or Fern in the materials reviewed, has not disclosed revenue or pricing after the beta, and has not said how many teams are using Lor. The company says it is still a small team and is hiring. What it does with $1 billion, and whether outside researchers can replicate its one-person model-building claim, will show whether the approach works beyond Mirendil's own walls.
How we reported this
DailyMech read Mirendil's Series A announcement and its October product post introducing Lor, and SiliconANGLE's report on the round. Excerpts of Sidley Austin's deal note and the Tech Startups roundup were consulted only for Sidley's role and the undisclosed valuation. Claims about training cost, timelines, agent-written code, parameter counts and Lor's capabilities are Mirendil's own, published directly or relayed by SiliconANGLE; DailyMech did not use Lor or test the models. DailyMech contacted no one for this article.
Published November 13, 2026. Facts checked against the cited sources by DailyMech's AI writer.
Sources
- Mirendil Raises $1B to Democratize Frontier AI R&D, Mirendil, Nov. 12, 2026 (fully read).
- Meet Lor, Mirendil, Oct. 28, 2026 (fully read).
- Mirendil raises $1B in funding to speed up AI model development, Maria Deutscher, SiliconANGLE, Nov. 13, 2026 (fully read).
- Sidley Represents Clearlake Capital Group in Mirendil's US$1 Billion Financing, Sidley Austin, Nov. 13, 2026 (excerpt read).
- Startup Funding News Today, November 13, 2026, Nickie Louise, Tech Startups, Nov. 13, 2026 (excerpt read).
DailyMech used AI tools to research and draft this article. Facts were checked against the cited sources by DailyMech's AI writer, not by a human editor.