Mirendil raises US$1B Series A to democratise frontier AI R&D

Dealroom describes Mirendil as a San Francisco AI lab founded by former Anthropic researchers[2]. Chief executive Behnam Neyshabur co-led Google DeepMind's Blueshift team[2]. Chief technology officer Harsh Mehta built the first version of Anthropic's internal AI research platform[2].

Dealroom puts five months between the seed and this Series A[2]. It says Mirendil launched in June with a US$200 million seed led by Andreessen Horowitz and Kleiner Perkins, at a valuation of around US$1 billion, and that Nvidia joined later[2]. Unite.AI dates that launch to 24 June 2026[3]. It says Mirendil described a founding team of 20 researchers and engineers from institutions including Anthropic, xAI, Google DeepMind and OpenAI, and that investment from Nvidia and others followed the seed[3].

Dealroom says Mirendil has not disclosed a valuation for the Series A[2]. It also says that in September, Bloomberg reported the company was in talks to raise up to US$1 billion at a US$5 billion valuation[2].

Mirendil says results it has shared over the past three weeks point to the same shift: when agents do the hands-on work, the size of a team stops limiting what can be built[1].

One researcher, working with agents, pretrained Mirendil-Base, a competitive coding model, from scratch in 3.5 months with less than US$10 million in compute, the company says[1][2]. Unite.AI reports that Mirendil's 21 October 2026 report counts that sub-US$10 million figure as all-in: research and small-scale experiments, the final training run, agent token spending and data processing[3].

In that run, Unite.AI reports, agents wrote more than 99% of the code, ran the experiments and diagnosed failures[3]. The researcher chose what to try and what to trust, and intervened when needed[3]. The experiment used publicly available code, web pages and public datasets, without purchasing or licensing additional data, and without distillation or synthetic data[3]. Unite.AI says Mirendil-Base sits in the 100 billion to 500 billion parameter range[3]. The larger Mirendil-Base-Large and Mirendil-Base-XL models were still in training when the report was published[3].

Mirendil says it then post-trained Mirendil-Base into Fern, which holds up to 10 million tokens of context, enough for days of an agent's work[1]. Unite.AI identifies the first model as Fern 0.1[3]. It says the model page describes Fern for high-volume agent work such as debugging training runs, reading complete agent traces and auditing training data, with a context window of up to 10 million tokens, roughly 23 MB or about half a million lines of code. Access is listed as coming soon[3].

Lor, in private beta, is how Mirendil says this way of working reaches other people: hundreds of agents writing code and running experiments for days at a time[1]. The company says it runs on Lor every day[1]. Unite.AI says Lor was introduced on 28 October 2026 as a cloud system whose agents can write code, run experiments, provision compute and operate on machines the user selects. An agent can divide a job across dozens of agents, with hundreds running across projects, and agents can continue for days or weeks and unblock themselves when something breaks[3]. The private beta is free, by invitation, on macOS, Linux and the web[3]. Windows, iOS and Android are listed as coming soon[3].

Mirendil says it is still a small team with a singular focus. People who want to make frontier AI research and development accessible can join, and anyone who wants to build with Mirendil can write to contact@mirendil.com[1].

Mirendil says the round lets it go further: one person can take on what used to need a team, a team what used to need an institution, and the lab can help accelerate science and technology more directly than before[1]. Unite.AI says the Series A connects agent-assisted model development, a platform for delegating research work, and cloud infrastructure for larger experiments[3]. It frames the central test as whether the one-researcher result becomes a repeatable workflow for scientists and engineers outside the company[3].

Unite.AI says Mirendil had already announced a multi-year Google Cloud partnership on 6 August 2026[3]. The partnership puts training, inference and AI research workloads on Google Cloud's AI Hypercomputer, with Google Cloud TPUs, full-stack Nvidia AI infrastructure, and managed training clusters co-designed across compute, storage, networking and control planes[3].

Dealroom calls a US$1 billion Series A rare even in AI, and places Mirendil among the best-funded of the new AI research labs, alongside Safe Superintelligence, Thinking Machines Lab and Recursive Superintelligence[2]. It says the round lands a week before Anthropic, where much of the founding team came from, is set to list on Nasdaq[2].