Mirendil, a San Francisco startup building agents that do the hands-on work of AI research, has raised $1 billion in a Series A round led by Kleiner Perkins and Clearlake Capital, the company announced on November 12. Google, Battery Ventures, Doerr Capital, Gil & Co, M13, Pear VC and S32 also took part. The size of the round, unusual for a first institutional raise, is a bet that agents can shrink the teams needed to build frontier models, a claim Mirendil backs with a coding model it says one researcher pretrained in 3.5 months for under $10 million.

Who is behind Mirendil

The company was founded last year by two researchers who worked at Google DeepMind and Anthropic. Chief executive Behnam Neyshabur is a co-author of Sharpness-Aware Minimization (SAM), a widely cited training technique that helps models generalize beyond the data they were trained on. Chief technology officer Harsh Mehta worked on pretraining at both labs, according to SiliconANGLE. Mirendil did not disclose its valuation and describes itself as a small team.

What Mirendil builds

The company has three products, all announced in the past three weeks:

  • Lor, now in private beta, is a platform where hundreds of agents write code and run experiments for days at a time. Mirendil says it uses Lor for its own research every day.
  • Mirendil-Base is a coding-focused base model in the 100 billion to 500 billion parameter range. Larger versions are in training: Mirendil-Base-Large (500 billion to 2 trillion parameters) was about 80% of the way through and Mirendil-Base-XL (2 to 10 trillion) about 19%, the company said in late October.
  • Fern is Mirendil-Base after a short round of post-training, with a context window of up to 10 million tokens. Mirendil says that is about 23 MB of code, or roughly half a million lines, enough to hold days of an agent’s work or a codebase the size of Megatron-LM at once.

The one-researcher pretraining claim

Mirendil’s central claim is that a single researcher, working with agents, pretrained Mirendil-Base from an empty repository. The company says agents wrote more than 99% of the code, ran experiments and diagnosed failures, while the researcher decided what to try and what to trust. The less-than-$10 million figure is described as all-in: the final run, the small-scale experiments before it, agents’ token spending and data processing. Mirendil says it used only public data, mainly GitHub code and web pages, with no distillation, synthetic data or licensed datasets.

Rows of black server racks with cabling in a data center room
Pretraining a base model still needs data center compute; Mirendil says its full Mirendil-Base effort cost less than $10 million. Photo: Carl Lender from Sunrise, USA / Wikimedia Commons, CC BY 2.0

These are the company’s own figures and have not been independently checked. Mirendil compares its models with others mainly on bits per byte, a measure of how well a model predicts held-out code, and says fuller long-context and agentic benchmark results are coming. It also corrected two charts in its Fern announcement after publication.

How it compares with the cost of a frontier team

To see why investors paid attention, compare the claim with how labs usually work. Open-model efforts typically involve dozens of specialists for architecture, data cleaning and cluster operations, and frontier labs employ hundreds. Labs themselves have started reporting that agents handle a growing share of their engineering: our report on METR’s audit of AI research acceleration at Anthropic covered one such measurement. Mirendil’s pitch is to sell that capability to outside teams rather than keep it inside one lab.

The round also fits a broader shift in funding toward the infrastructure that runs agents, after LangChain raised $300 million at a $3 billion valuation the day after Mirendil’s announcement.

Who could use it

If Lor works as described, the first customers would be research groups, model startups and companies that train their own models but cannot hire large pretraining teams. Mirendil says it wants to put frontier AI research “within reach of every scientist and engineer”. The open question is how much human judgment the process still needs: Mirendil’s own account credits the researcher with steering every major decision.

What to watch next

  • Independent benchmark results for Fern and Mirendil-Base on standard coding and agent tests.
  • When Lor leaves private beta, and how it is priced.
  • Results from Mirendil-Base-Large and Mirendil-Base-XL once training finishes.
  • Whether other labs publish comparable small-team or agent-led pretraining runs.

Key takeaways

  • Mirendil raised $1 billion in a Series A led by Kleiner Perkins and Clearlake Capital, with Google among the investors.
  • Its Lor platform runs hundreds of AI agents that write code and run experiments for days.
  • The company says one researcher pretrained its Mirendil-Base model in 3.5 months for under $10 million.
  • Its Fern model supports a context window of up to 10 million tokens.
  • All performance and cost figures so far come from Mirendil and await outside verification.

Frequently asked questions

What is Mirendil?

Mirendil is a San Francisco AI startup founded by former Google DeepMind and Anthropic researchers. It builds Lor, a platform of research agents, and its own coding models, Mirendil-Base and Fern.

How much money has Mirendil raised?

Mirendil announced a $1 billion Series A on November 12, 2026, led by Kleiner Perkins and Clearlake Capital. It did not disclose its valuation.

What is Mirendil Lor?

Lor is Mirendil’s platform for automating AI research and engineering, in which hundreds of agents write code and run experiments for days. It was in private beta at the time of the funding announcement.