Three Weeks After Its First Model, Mirendil Lands a $1B Series A to Put a Whole AI Lab Behind One Researcher

Mirendil, a young AI company building tools that let agents do the hands-on work of AI research, has raised $1 billion in a Series A. The company announced the round on November 12, 2026. Kleiner Perkins and Clearlake Capital led it, with participation from Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32, among other new and existing investors. Mirendil did not disclose a valuation.

The size of the round stands out because of how recently Mirendil started showing its work. Its first public results arrived only about three weeks earlier, in a run of posts between October 21 and October 28 covering a base model, a long-context model and an agent platform.

The thesis: team size stops being the limit

Mirendil’s stated goal is to “democratize frontier AI R&D” by putting it within reach of every scientist and engineer. Its argument is that when agents do the hands-on engineering, the number of people on a team stops capping what it can build. In the funding post, the company puts it plainly: one person should be able to take on what used to need a team, and a team what used to need an institution.

The company says it is still a small team with a single focus, and that the new money lets it push further toward that goal.

What it has shown so far

The evidence Mirendil points to is its own model. On October 21, it said a single researcher working with agents pretrained Mirendil-Base, a coding model, from scratch in 3.5 months for less than $10 million in compute. Mirendil said agents wrote more than 99% of the code, and that training used public data only, without distillation. At the time it described a family in progress: a Base model in the 100–500B range, a Large model of 0.5–2T parameters about 80% trained, and an XL model of 2–10T parameters about 19% trained.

On bits-per-byte across seven external code repositories, Mirendil reported 0.222 for Mirendil-Base, behind Kimi-K3 at 0.193 but ahead of GLM-5.3 (0.242), DeepSeek-V4-Pro-Base (0.243) and Nemotron-550B-Base (0.247). Those are company-run measurements.

Mirendil then post-trained the base into Fern, launched on October 26 in early access. Fern holds up to 10 million tokens of context, which the company says is enough to keep days of an agent’s work in view at once.

Lor is the product

The funding is mainly a bet on Lor, which Mirendil introduced on October 28 and says it uses every day internally. Lor is built to run agents that write code, run experiments and iterate on results for days or weeks, hundreds at a time.

Several design choices separate it from a typical coding assistant:

• Agents provision their own compute. They request nodes, set them up, run the work and release them afterwards, with no cloud account for the user to configure.
• The agent and the machine are separate. Agents run in Mirendil’s cloud but send commands to whichever machine the user picks: a Lor-managed computer, a laptop or another always-on box. Closing a laptop does not stop the work, and machines can be removed at any time.
• Work fans out. One job can split across dozens of agents. In one example, Mirendil says 102 agents built environments from 100 PyTorch pull requests, averaging eight minutes each. In another, agents found and fixed security bugs in LiteLLM overnight.

Lor is not limited to Mirendil’s own models. Its model list includes Anthropic’s Opus 5.5 and Fable 5.1, OpenAI’s GPT-6 Astra and GPT-6.1 Sol, Google’s Gemini 4 Argon, xAI’s Grok 4.7, Meta’s Muse Spark 1.3, and open models from Z.ai, Moonshot AI and DeepSeek, alongside Fern.

Invite-only beta, no spending breakdown yet

Lor is in private beta by invitation, with a waitlist. It runs on macOS, Linux and the web, with Windows, iOS and Android listed as coming soon, and it is free during the beta. Fern remains in early access with no published pricing or open weights.

Mirendil has not said how it will split the $1 billion between compute, model training and hiring, and it has not published customer numbers. For now, the round is a large early wager that agent-run research can shrink the cost and headcount of building frontier models.