Mirendil raises $200M seed led by a16z and Kleiner Perkins to bring frontier AI research outside the big labs

What's the deal? Mirendil, a startup building a platform that lets engineers and researchers do frontier AI work outside the big labs, has raised a $200 million seed round at a roughly $1 billion valuation. Andreessen Horowitz (a16z) and Kleiner Perkins co-led the round, with participation from Nvidia and others.

Mirendil trains frontier models that specialise in AI research and development, then wraps a product around them. The system works like a coding agent built for AI tasks: it controls its own GPUs and loops over research and engineering problems autonomously, with less human intervention over time.

Why now? Scaling laws have concentrated AI talent and resources in a few labs, since training large models costs tens of billions of dollars and needs hundreds of thousands of GPUs. a16z argues the technology's full potential will not arrive until it reaches builders outside those labs — the range of problems language models can solve is too vast for a handful of companies to tackle alone. Most major labs have built similar platforms internally, but they have little incentive to release them to competitors.

What could go wrong? This is not a simple product to build. a16z calls it a systems problem, where backend infrastructure, agent harness, evals, post-training, and even pre-training strategy must be designed together. The training data also needs to cover the full research loop: proposing experiments, writing and running code, interpreting results, debugging failures, and deciding what to try next. Mirendil's product will likely serve engineers and researchers first, with the goal of supporting less technical users like scientists later.

The signal: a16z calls the approach "vibe research," and the bet is that the domain expertise living outside the labs unlocks impact across fields like materials science, drug discovery, and enterprise software — a thesis reflected in Mirendil's own positioning around biology and materials science research. The firm points to CursorDealroom has a profile for this one. Try Dealroom →'s path from third-party models to pre-training its own as evidence that the modelling work done outside the big labs can compound in the same way as their centralised efforts.

Read more: Andreessen Horowitz