Mirendil

Mirendil

Mirendil is a frontier AI research lab whose product is the research loop itself. Mirendil is building systems that plan and run machine learning experiments autonomously — described by its investors as a coding agent built for AI tasks, controlling its own GPUs and iterating over research and engineering problems with progressively less human involvement. The stated goal is to put frontier-scale R&D within reach of universities, institutes and companies that cannot staff a large lab, and to point the same loop at adjacent fields such as drug discovery, chemistry and robotics.

Production credibility: Mirendil raised a $200M seed at roughly a $1B valuation, led by Andreessen Horowitz and Kleiner Perkins with NVIDIA participating — one of the largest seed rounds in the sector. It was founded by Behnam Neyshabur and Harsh Mehta, who left Anthropic in December 2025, and the team includes around twenty researchers drawn from Anthropic, xAI, DeepMind and OpenAI. The company's own site is deliberately sparse: it states the mission — democratising frontier AI R&D to accelerate science and technology — without documenting the technology, so the public detail comes from investors and press rather than product pages.

Key Features

• Systems that plan and run machine learning experiments with limited supervision
• Agent loop that controls its own compute rather than proposing runs to a human
• Aimed at model training and product building, not just paper writing
• Positioned for research groups outside the largest frontier labs
• Intended to extend to drug discovery, chemistry and robotics
• Team drawn from Anthropic, xAI, DeepMind and OpenAI

Ideal Use Case

Research organisations and companies that want frontier-scale experimentation without frontier-scale headcount: universities, national institutes, and teams whose domain expertise is in science rather than in ML infrastructure. At this stage it is a company to track rather than a product to buy, so the practical use case is evaluation and early access conversations.

How Mirendil differentiates

Thinking Machines, Periodic Labs and Lila Sciences are also research-first companies rather than tool vendors, but they aim their systems at a domain — science, chemistry, materials. Mirendil aims its system at AI research itself, which is a narrower claim and a more recursive one: the output is meant to be better models, which then improve the loop. The caveat is transparency. There is no public product, no published benchmark and no pricing; the evidence available today is the team, the investors and the size of the round.

FAQ

Q: What does Mirendil do? A: Mirendil builds AI systems that carry out AI research: planning and running machine learning experiments, controlling their own compute, and iterating with decreasing human intervention.

Q: Who founded Mirendil? A: Behnam Neyshabur and Harsh Mehta, who left Anthropic in December 2025. The team includes around twenty researchers from Anthropic, xAI, DeepMind and OpenAI.

Q: How much has Mirendil raised? A: Mirendil raised a $200M seed round at roughly a $1B valuation, led by Andreessen Horowitz and Kleiner Perkins, with NVIDIA participating.

Q: Can I use Mirendil today? A: Not as a self-serve product. The company publishes a mission statement rather than product documentation or pricing, so access comes through direct contact.

Q: How is Mirendil different from other AI labs? A: Most labs build models for a domain. Mirendil is building a system that does the research work itself, with the stated aim of making frontier-scale R&D available to institutions outside the largest labs.

tl;dr

Mirendil is a frontier AI lab founded by ex-Anthropic researchers, building systems that run AI research autonomously and control their own GPUs. It raised a $200M seed at roughly $1B led by a16z and Kleiner Perkins, with NVIDIA participating. No public product yet.