Mirendil raises $1 billion to industrialize AI research
On November 12, 2026, Mirendil announced a $1 billion funding round led by Kleiner Perkins and Clearlake Capital. The startup aims to make cutting-edge R&D more accessible through models, agents, and a delegated work platform.
On November 12, 2026, Mirendil announced a $1 billion Series A funding round led by Kleiner Perkins and Clearlake Capital to make cutting-edge AI research and development more accessible to scientists and engineers. The startup also aims to demonstrate that a small group can accomplish tasks previously reserved for entire teams or even institutions. Its bet relies on foundation models, software agents, and a cloud infrastructure designed for longer research loops. The market is mainly watching one simple question: is this model reproducible, or only spectacular on paper?
A massive funding round for a very specific goal
Mirendil gathered a funding round including Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC, and S32, among other new and historic investors. The company is not just financing a product team; it seeks to build a complete chain, from model training to agent-assisted research work execution. In this context, the size of the check matters less than the industrial thesis behind the funding. And this thesis is clear: to reduce the cost, time, and number of humans needed to produce useful research.
The startup was founded by 20 researchers and engineers from Anthropic, xAI, Google DeepMind, and OpenAI, according to its announcement on June 24, 2026. It had already raised $200 million in seed funding, a round led by Andreessen Horowitz and Kleiner Perkins, with participation from NVIDIA and other investors. For a young company, that is quite a lot of capital for a single idea. But the idea is far from trivial: if it works, it directly impacts how the next generation of models is built.
The technical test that served as proof
Mirendil says it published its first model results on October 21, 2026, and this is where the story becomes more concrete. According to the company, a single researcher, aided by pre-trained AI agents, trained Mirendil-Base, a code model, from scratch in 3.5 months with less than $10 million in compute. This figure includes research, small tests, final training runs, tokens consumed by agents, and data processing.
The company claims that agents wrote more than 99% of the code, launched experiments, and diagnosed failures, while the researcher decided what to try and what to keep. The experiment relied on public code, web pages, and public datasets, without additional data purchases or licenses, and without distillation or synthetic data. Mirendil-Base is estimated to have between 100 billion and 500 billion parameters [to be verified], while the larger models Mirendil-Base-Large and Mirendil-Base-XL were still in training at the time of the report.
Fern and Lor, two building blocks to delegate heavy work
After this first milestone, Mirendil post-trained Mirendil-Base to create Fern 0.1, presented as its first model. The product page describes a system designed for high-volume agentic tasks, such as debugging training runs, reading full agent traces, or auditing training data. Its architecture, training, and service support a context window of up to 10 million tokens, about 23 MB or half a million lines of code. In other words, enough to keep a full day’s work of an agent in memory as a single block.
The other pillar is called Lor. Launched on October 28, 2026, this cloud system allows agents to write code, run experiments, provision compute, and operate on machines chosen by the user. Mirendil states that an agent can split a task among dozens of others, with hundreds of active instances across multiple projects, and that these agents can continue for days or weeks while self-unblocking when something breaks. Yes, this is the kind of tool that gives project managers cold sweats and ideas to labs.
Access to Lor is by invitation only, in private beta. It is free during this phase on macOS, Linux, and the web, while Windows, iOS, and Android are announced as coming soon. Users can track each agent’s progress, stop it, and remove machines from the system.
The crux of the matter remains infrastructure
Mirendil also announced on August 6, 2026, a multi-year partnership with Google Cloud. The agreement places training, inference, and AI research workloads on Google Cloud’s AI Hypercomputer, relying on Google Cloud TPUs and NVIDIA’s full AI infrastructure. It also includes managed training clusters co-designed for compute, storage, networking, and control planes to support broader research loops.
This dimension matters as much as the product itself. Without stable compute capacity, any talk of automating research remains a stage demonstration. With it, Mirendil attempts to connect three layers: model creation, agent orchestration, and cloud scaling. The real question, less glamorous and more decisive, remains: can a successful internal experiment become a reliable workflow for external teams, with the same level of control and cost?
Key points
- $1 billion raised in Series A on November 12, 2026.
- Kleiner Perkins and Clearlake Capital led the round.
- One researcher trained Mirendil-Base in 3.5 months.
- Less than $10 million in compute announced.
- Lor is in private beta by invitation, free during this phase.
- Google Cloud provides the infrastructure for the partnership announced in August 2026.
By the numbers
- $1 billion — Series A announced by Mirendil, November 12, 2026.
- 3.5 months — claimed training duration for Mirendil-Base.
- Less than $10 million — total compute cost announced for the experiment.
- 10 million tokens — maximum context window claimed for Fern 0.1.
- 20 researchers and engineers — founding team announced in June 2026.