Mirendil has raised $1bn in a Series A to build tools for AI research, and says one researcher used AI agents to train a competitive coding model from scratch in three and a half months for less than $10m in compute. The funding is substantial; the more intriguing claim is that agents could let much smaller teams tackle work once reserved for larger labs.
Mirendil Watch analysis
What happened
Mirendil announced the round on 12 November, led by Kleiner Perkins and Clearlake Capital, with Google and several other investors participating. The company says it used AI agents to train Mirendil-Base, then post-trained the model into Fern, which it describes as having a 10-million-token context window.
It also says Lor, currently in private beta, coordinates hundreds of agents to write code and run experiments over several days. Mirendil says it uses Lor itself. These results and product descriptions come from the company’s announcement, not an independent evaluation.
Why it matters
The pitch is that agents could change who can do AI research, not merely make existing teams a little faster. If one researcher can train a competitive model at the reported cost and timescale, that would be a striking example of AI tools lowering the labour and compute needed for model development.
But “competitive” is Mirendil’s characterisation, and the announcement does not give benchmark results or enough detail to assess the model’s performance. The claims are worth watching, not yet a reason to declare the research lab obsolete. It is a rather large bet on a small-team future, in both senses of the word.
Our read
Mirendil has put forward a clear thesis and a concrete example, alongside enough funding to pursue it. The useful test is whether the company publishes evaluations and shows that its agent-driven research can be repeated beyond one reported project. For now, the announcement makes the case; independent results would make it convincing.
What to watch
- Whether Mirendil publishes benchmark results and training details for Mirendil-Base and Fern.
- When Lor leaves private beta and what access will cost.
- Whether other researchers can reproduce the reported training time and compute budget.
Discussion spark: If one researcher can train a competitive model with agents for under $10m, does that meaningfully open up AI research, or do compute and evaluation still keep the field in the hands of large organisations?
Sources and evidence
- Source update (12 November 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Mirendil. Not affiliated with or endorsed by Mirendil; this is not an official account.