> ## Content Index
> Fetch the complete content index at: https://venture-block.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Ex-Anthropic researchers' Mirendil raises $1B Series A led by Kleiner Perkins and Clearlake
- URL: https://venture-block.com/mirendil-1b-series-a-kleiner-clearlake-ai-research-agents/
- Published: 2026-11-13T14:05:23.000Z
- Updated: 2026-11-13T14:05:23.000Z
- Description: The former Anthropic team behind Mirendil raised $1 billion less than five months after its $200 million seed, betting that agents can let one researcher pretrain models that once took a full team.
- Author: Venture Block
- Tags: Funding, AI

Mirendil, the San Francisco AI lab started by former Anthropic and Google DeepMind researchers, [has raised $1 billion in a Series A](https://mirendil.com/news/series-a/?ref=venture-block.com) led by Kleiner Perkins and Clearlake Capital. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also joined the round, which the company announced on Nov. 12 along with other unnamed new and existing investors.

The company did not disclose a valuation. In September, [Bloomberg reported](https://news.bloomberglaw.com/artificial-intelligence/ex-anthropic-staffers-ai-startup-in-talks-for-5-billion-value?ref=venture-block.com) that Mirendil was in talks to raise up to $1 billion at a $5 billion valuation including the new money, with Kleiner in talks to lead. Mirendil has not confirmed whether the final terms match.

The round lands less than five months after Mirendil came out of stealth. On June 24 it [announced a $200 million seed round](https://mirendil.com/news/announcing-mirendil/?ref=venture-block.com) led by Andreessen Horowitz and Kleiner Perkins, followed by an investment from Nvidia. [SiliconANGLE reported](https://siliconangle.com/2026/06/25/mirendil-raises-200m-speed-scientific-research-ai/?ref=venture-block.com) that seed valued the company at $1 billion.

## Agents doing the lab work

Mirendil's pitch is that AI agents can now do most of the hands-on engineering behind frontier AI research, so a single researcher can run a project that once needed a full team. Chief executive Behnam Neyshabur previously worked on AI for science at Google and Anthropic, and chief technology officer Harsh Mehta built the first version of Anthropic's internal automated research platform, according to [Andreessen Horowitz's investment note](https://a16z.com/announcement/investing-in-mirendil/?ref=venture-block.com). The company says its founding team of 20 came from Anthropic, xAI, Google DeepMind and OpenAI.

Its main evidence so far is Mirendil-Base, a coding-focused model the company [says one researcher pretrained from scratch](https://mirendil.com/news/pretraining-at-scale-without-a-pretraining-team/?ref=venture-block.com) in 3.5 months for less than $10 million in compute. Mirendil describes that figure as all-in, covering research, small experiments, the final training run, agent token spend and data processing. It says agents wrote more than 99% of the code, ran experiments and diagnosed failures, while the researcher chose what to try and when to step in. The model was trained only on public code, web pages and datasets, with no distillation and no purchased data.

Mirendil-Base sits in the 100 billion to 500 billion parameter range. On the company's own test, which measures how well models predict code in seven external repositories published after training, it scored an average of 0.222 bits per byte, second to Moonshot's Kimi-K3 at 0.193 and ahead of GLM-5.3, Nemotron-550B-Base and DeepSeek-V4-Pro-Base. Lower is better. That is a prediction-loss benchmark chosen and run by Mirendil, not a coding-task score, and no outside group has reproduced it, as [Tech Times noted](https://www.techtimes.com/articles/330614/20261113/ai-lab-mirendil-raises-1b-series-release-names-no-valuation-cites-only-company-tests.htm?ref=venture-block.com).

Two larger models are still training. Mirendil said on Oct. 21 that Mirendil-Base-Large (500 billion to 2 trillion parameters) was about 80% through training and Mirendil-Base-XL (2 trillion to 10 trillion) about 19% through. The sub-$10 million cost figure covers only the smallest model.

## Fern and Lor

A short round of post-training turned Mirendil-Base into [Fern](https://mirendil.com/news/introducing-fern/?ref=venture-block.com), which the company says supports up to 10 million tokens of context, roughly half a million lines of code. Mirendil pitches it for the routine work of building AI: auditing training data, debugging a run from its logs and reading long agent traces end to end. Fern's model page lists access as coming soon.

The product customers can try is [Lor](https://mirendil.com/news/meet-lor/?ref=venture-block.com), an agent platform launched in private beta on Oct. 28\. Lor runs hundreds of agents in parallel for days, spinning up GPUs when they need them and releasing them afterward. Alongside Fern it runs models from Anthropic, OpenAI, Google, xAI, Meta, Moonshot, Z.ai and DeepSeek. It is invite-only, free during the beta, and available on macOS, Linux and the web. Mirendil says it runs its own work on Lor every day.

## Compute and competition

A large share of the money is likely to go to compute. In August, Neyshabur [told TechCrunch](https://techcrunch.com/2026/08/06/exclusive-mirendil-inks-100m-google-cloud-deal-to-scale-self-improving-ai/?ref=venture-block.com) that a multiyear Google Cloud deal giving Mirendil access to TPUs and Nvidia GPUs was worth upward of $100 million. Google is now also an investor.

Mirendil is one of several well-funded efforts built around AI that speeds up AI research. TechCrunch has named Recursive Superintelligence and Ricursive Intelligence among startups chasing the same goal, and the big labs are working on it in-house: Anthropic's amended IPO filing [warned that its AI may soon sharply speed up its own research](https://venture-block.com/anthropic-s1a-ai-accelerated-research-q3-revenue/). Mirendil's difference is that it wants to sell the tooling to outside scientists and engineers rather than keep it inside one lab. Its benchmark rival Kimi comes from Moonshot, which [plans an open-weights release of Kimi K3.1 Pro](https://venture-block.com/moonshot-kimi-k3-1-pro-open-weights/), and AI evaluation is drawing big checks of its own, as [Mercor's $750 million round](https://venture-block.com/mercor-750m-series-d-20b-valuation-enterprise-evals/) showed this week.

Mirendil says it remains a small team and is hiring. Its next public markers are post-training results for Mirendil-Base, broader access to Fern, and updates on the Large and XL models still in training.