Mirendil raises $1 billion Series A after $200 million seed round

Mirendil raised $1 billion in a Series A led by Kleiner Perkins and Clearlake Capital, following a $200 million seed round announced in June, Unite.AI reported. The San Francisco AI research startup announced the financing on 12 November 2026 to develop models and agents intended to let scientists and engineers undertake research work with smaller teams.

Kleiner Perkins returns after $200 million seed

Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also participated in the Series A. Mirendil announced its $200 million seed round on 24 June 2026, led by Andreessen Horowitz and Kleiner Perkins. Nvidia subsequently invested, according to Unite.AI.

Mirendil was established in late 2025 by former Anthropic employees Behnam Neyshabur and Harsh Mehta, Shayan Salehian, who previously worked at xAI, and Tara Rezaei, who had interned at OpenAI, Stockwhiz reported. It identifies Neyshabur as chief executive and Mehta as chief technology officer. At its June launch, Mirendil described a founding group of 20 researchers and engineers drawn from organisations including Anthropic, xAI, Google DeepMind and OpenAI.

The company’s financing ambitions had surfaced before the Series A: AI Affairs reported that Mirendil was seeking a $5 billion valuation, five times its seed value. The new financing gives Mirendil capital for its model, agent and cloud-infrastructure plans. Kleiner Perkins has backed both announced rounds, while Clearlake Capital joins as a lead investor in the Series A.

Mirendil-Base training cost less than $10 million

Mirendil said in its report of 21 October 2026 that a single researcher directed AI agents to train its Mirendil-Base coding model from the outset over 3.5 months. The company put compute spending at less than $10 million, including preliminary research and experiments, the final training run, agent tokens and data processing. It placed the model in the 100 billion to 500 billion parameter range.

Mirendil says agents produced more than 99% of the code, conducted experiments and investigated failures. The researcher selected approaches, assessed the agents’ work and intervened when needed. According to the company, the training used publicly available code, web pages and datasets, without additional purchased or licensed data, distillation or synthetic data.

Mirendil said its larger Mirendil-Base-Large and Mirendil-Base-XL models were still training when it published those results. It subsequently post-trained Mirendil-Base into Fern 0.1, a model intended for tasks such as examining agent traces, debugging training runs and auditing training data. The company says Fern supports a context window of up to 10 million tokens, enough to retain extended records of an agent’s work. Access was listed as coming soon.

Mirendil’s Google Cloud deal covers AI workloads

Mirendil introduced Lor on 28 October 2026 as a cloud platform for delegating research tasks. Its agents can write code, run experiments and provision compute on machines selected by users. The company says one agent can split work among dozens of others, with hundreds operating across projects, and that agents can continue working for days or weeks.

The platform is available by invitation in a private beta and is free during that period on macOS, Linux and the web. Mirendil lists Windows, iOS and Android versions as coming soon. Users can monitor an agent’s progress, stop it and remove machines from the system.

Mirendil announced a multi-year Google Cloud partnership on 6 August 2026 covering training, inference and AI research workloads. The arrangement uses Google Cloud’s AI Hypercomputer, including Google Cloud TPUs and Nvidia infrastructure. It also covers managed training clusters designed across compute, storage, networking and the systems that control them.