AI News

Mirendil, founded by former Anthropic employees, completes a $1 billion Series A funding round with Google participating

Phoenix TechnologyUpdated Nov 16 22:56

AI summary

U.S. large-model unicorn Mirendil announced on November 13 that it had completed a $1 billion Series A funding round, co-led by Kleiner Perkins and Clearlake Capital, with participation from Google, Nvidia, and others; multiple media outlets, citing internal a16z information, reported its valuation at approximately $5 billion. Including its $200 million seed round in June this year, Mirendil has raised a total of $1.2 billion. The company was founded by former Anthropic employees Behnam Neyshabur and Harsh Mehta together with former xAI researcher Shayan Salehian and former OpenAI student researcher Tara Rezaei. On October 21, it released its first foundation model, Mirendil-Base; the company officially stated that one employee, with the help of agents, completed pretraining from scratch in three and a half months, with total compute costs below $10 million.

Why read this

AI score 69

Beyond the funding and the founding team's backgrounds, the original article also provides the cost and time required for one person to complete pretraining with the help of agents, which can serve as a sample for observing changes in the barriers to AI R&D.

Topics

Tags

Body · AI translation

Rakes in 8 Billion Yuan! Former Anthropic Employees Bet on RSI, with Google and Nvidia Both Investing

Zhidongxi Download the App for Exclusive Early Access November 16, 2026, 21:55:11 From Beijing

Zhidongxi reported on November 16 that on November 13, the U.S. large-model unicorn Mirendil officially announced the completion of a $1 billion (approximately 6.672 billion yuan) Series A funding round. The company did not disclose the valuation for this round; multiple media outlets, citing internal a16z information, reported its valuation at approximately $5 billion (approximately 33.359 billion yuan).

This round was co-led by Kleiner Perkins and Clearlake Capital, with participation from Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC, S32, and others.

Including the seed round in June this year, Mirendil has raised a cumulative total of $1.2 billion (approximately 8.006 billion yuan).

▲Mirendil officially announces funding (Source: X)

Mirendil emerged from stealth in June 2026. It was founded by former Anthropic employees Behnam Neyshabur and Harsh Mehta, together with former xAI researcher Shayan Salehian and former OpenAI student researcher Tara Rezaei, and is headquartered in San Francisco, USA.

In a previous interview, Behnam Neyshabur said: "The product we are building is somewhat like AI for AI for science, rather than AI for science." In other words, Mirendil provides a tool to help researchers develop models themselves.

On October 21, 2026, Mirendil released its first foundation model, Mirendil-Base. The company said the model was pre-trained from scratch by one employee with the help of agents in just three and a half months, with total compute costs of less than $10 million (approximately 66.7175 million yuan).

Could this company, founded by former Anthropic employees, have already achieved RSI (Recursive Self-Improvement)?

1. Two funding rounds completed 141 days apart, with a compute deal worth over $100 million signed with Google Cloud

Since its founding, Mirendil has completed two consecutive funding rounds totaling $1.2 billion (approximately 8.006 billion yuan). On June 25, 2026, the company announced the completion of a $200 million (approximately 1.334 billion yuan) seed round led by a16z and Kleiner Perkins, with participation from Nvidia and other institutions. According to The Wall Street Journal, Mirendil was valued at $1 billion (approximately 6.672 billion yuan) at that time.

In terms of compute deployment, in August 2026, Mirendil signed a multi-year strategic cooperation agreement with Google Cloud.

According to information disclosed by both parties, the cooperation is worth more than $100 million (approximately 667 million yuan), mainly covering two types of compute resources: Google TPUs and NVIDIA GPUs. The agreement provides stable infrastructure support for Mirendil's subsequent large-scale model pre-training and agent experiments.

▲Mirendil officially announces cooperation (Source: X)

2. Pre-training completed in 3.5 months, with agents doing 99% of the programming work

On October 21, 2026, Mirendil released its first foundation model, Mirendil-Base, with a parameter scale of approximately 100 billion to 500 billion.

The company claims that the model was pre-trained from scratch by a single researcher with agent assistance, taking only 3.5 months, with total compute costs below $10 million (approximately 66.7175 million yuan), including experiments, agent token consumption, and data processing.

Among this, agents completed more than 99% of the programming work, handling infrastructure setup, algorithm selection, data collection, and iteration; humans were mainly responsible for directional decisions and evaluating results.

Subsequently, the company conducted post-training based on Mirendil-Base and further launched Fern 0.1. This model has a parameter scale of approximately 0.1 trillion to 0.5 trillion and supports a context window of up to approximately 10 million tokens.

The model is mainly aimed at high-capacity tasks such as debugging training logs, analyzing complete agent trajectories, and auditing pre-training data. It is currently in the early access application stage.

At the same time, Mirendil previewed that larger-parameter versions are being advanced in parallel: Mirendil-Base-Large (approximately 500 billion to 2 trillion parameters) is about 80% complete, and Mirendil-Base-XL (approximately 2 trillion to 10 trillion parameters) is about 20% complete.

▲Preview of model progress

Its core product, the multi-agent R&D platform Lor, entered private testing on October 28, 2026. Mirendil revealed that its company's internal daily R&D is already fully running on Lor.

The platform can simultaneously schedule hundreds of agents to write code, run experiments, and automatically request and release compute resources, supporting continuous work for days or even weeks.

▲In one run, the agents in Lor built environments based on 100 PyTorch pull requests: 102 agents, averaging 8 minutes each

▲Agents allocate compute resources on demand and have full control over every virtual machine and every node they launch

▲In one overnight run, the agents in Lor discovered a security vulnerability in LiteLLM and completed the fix before dawn

In Lor, agents are separated from the devices they work on. The agents decide what to run, and Lor sends each command to the device selected by the user. Moreover, agents only operate on machines that have been added, and users can remove a device at any time.

▲Agents separated from devices

Currently, Lor supports macOS, Linux, and the web, with Windows, iOS, and Android versions planned. The platform is compatible with open-source models such as Kimi and DeepSeek, as well as frontier closed-source models such as Claude, GPT, Gemini, and Grok, while also supporting Mirendil's own models.

▲Supported models

III. A founding team from Anthropic, xAI, and OpenAI, covering model reasoning, automated R&D, and engineering implementation

Mirendil was founded by four co-founders, whose core members mostly come from Anthropic, Google DeepMind, xAI, and OpenAI, each having accumulated direct experience in model reasoning, automated R&D, engineering implementation, and early-stage research.

Co-founder and CEO Behnam Neyshabur holds bachelor's and master's degrees from Sharif University of Technology and a PhD in computer science from Toyota Technological Institute at Chicago. He then studied under Yann LeCun, conducting postdoctoral research at New York University and the Institute for Advanced Study in Princeton.

From 2019 to 2024, Behnam served as a senior research scientist at Google DeepMind and co-led the Blueshift team, focusing on improving the reasoning capabilities of foundation models. The team directly participated in the development of Minerva, the first language model to perform outstandingly on math problems, and was responsible for building the reasoning capabilities of early Gemini models; subsequent work also contributed to reasoning improvements in Gemini 1.5 and Gemma 2.

From late 2024 to December 2025, Behnam co-led the Discovery team at Anthropic, with the goal of building "AI scientists/engineers" capable of long-horizon reasoning and scientific tasks.

In research, he is a co-author of the ICLR 2021 paper "Sharpness-Aware Minimization" (the SAM optimizer), a technique widely used to improve the generalization ability of deep learning models; his research has been cited more than 50,000 times on Google.

▲Behnam Neyshabur (Source: Github)

Co-founder and CTO Harsh Mehta was Behnam's colleague from Google to Anthropic. He graduated from the Indian Institute of Technology Guwahati and previously worked at Google Search and Google DeepMind for nearly 10 years, participating in the development of the Gemini series of models and publishing papers in areas such as long-sequence modeling, state space models, and learning rate scheduling.

From 2024 to December 2025, Harsh served as a senior research scientist at Anthropic, leading the construction of the first version of the company's internal automated AI research platform (autoresearch). The platform aims to use models themselves to accelerate pretraining and experiment iteration, and is the direct predecessor of Mirendil's current technical direction.

▲Group photo of the Mirendil founding team, with Harsh Mehta on the far right (Source: wsj)

Co-founder Shayan Salehian received his bachelor's degree from Sharif University of Technology and studied for his master's at the University of California, Riverside.

From 2018 to 2024, he successively served as a machine learning engineer at Twitter and X, participating in timeline recommendation, model serving, and large-scale systems engineering.

From October 2024 to early 2026, Shayan joined xAI, participating in work related to post-training, reasoning, memory systems, coding infrastructure, and agents for the Grok series of models.

▲Shayan Salehian (Source: LinkedIn)

Co-founder Tara Rezaei (full name Tara Rezaei Kheirkhah) is the youngest member of the team. In 2021, she represented Iran at the International Olympiad on Astronomy and Astrophysics (IOAA) and won a silver medal, and subsequently entered MIT to study electrical engineering and computer science. During her studies, she served as a student researcher at OpenAI, with research involving model interpretability and program synthesis.

▲Tara Rezaei (Source: Tara's personal homepage)

Conclusion: Using Agents to Redefine the Barrier to AI R&D, Capital Bets on RSI

Today, frontier AI capabilities are highly concentrated in a few laboratories with massive compute, top talent, and a complete experimental loop. Mirendil attempts to break this structure with Agents, compressing pretraining and experiment iteration that originally required dozens of people and months or even years into a scale of "one person, several months, and tens of millions of dollars in compute."

Specialized laboratories in fields such as drug discovery, materials science, and robotics will have the opportunity to carry out near-frontier model iteration without building a complete pretraining team, which is likely to spawn a batch of "domain-specific small laboratories." Capital has bet on it twice within half a year, perhaps precisely because of this possibility.

But Mirendil's most impressive results so far are still internal company experiments. How does it perform externally? In an era of rapid iteration of frontier large models, how to build technical barriers? These are questions the startup team urgently needs to answer.

Sources: The Wall Street Journal, Mirendil official website, Google Cloud, LinkedIn

"Special statement: The above content (including videos, pictures, or audio if any) is uploaded and published by users of "Dafeng Hao," a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."

Source

Phoenix Technology