2026-11-17 18:06
Introduction: 99% of code written by agents, Lor orchestrates hundreds of Agents for R&D. SmartThings reported on November 16, 2026. On November 13, U.S.-based AI unicorn Mirendil officially closed a $1 billion Series A round (approximately RMB 6.672 billion).
SmartThings reported on November 16, on November 13, U.S.-based large model unicorn Mirendil announced the closure of a $1 billion Series A funding round (approximately RMB 6.672 billion). The company did not disclose valuation in this round; however, multiple media outlets cited internal information from a16z suggesting its valuation was around $5 billion (approximately RMB 33.359 billion).
This round was jointly 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 Series A round in June, Mirendil has now raised a total of $1.2 billion (approximately RMB 8.006 billion).

▲ Mirendil’s official funding announcement (Source: X)
Mirendil was unveiled from stealth mode in June 2026, founded by Behnam Neyshabur (former Anthropic employee) and Harsh Mehta (former xAI researcher), alongside Shayan Salehian (former xAI researcher) and Tara Rezaei (former OpenAI student researcher). The company is headquartered in San Francisco, USA.
In prior interviews, Behnam Neyshabur stated: “The product we are building is more like an AI for AI in scientific research (AI for AI for science), rather than AI for science (AI for science).” In other words, Mirendil provides a tool to help researchers develop their own models independently.
On October 21, 2026, Mirendil released its first foundational model, Mirendil-Base. According to official claims, this model was trained from scratch by a single researcher assisted by AI agents, completed in just 3.5 months, with total compute cost under $10 million (approximately RMB 6.67175 million), including experimental overhead, agent token consumption, and data processing expenses.
Could this company founded by former Anthropic employees have already achieved RSI (Recursive Self-Improvement)?
Since its founding, Mirendil has successfully completed two fundraising rounds totaling $1.2 billion (approximately RMB 8.006 billion). On June 25, 2026, the company announced a $200 million seed round (approximately RMB 1.334 billion), co-led by a16z and Kleiner Perkins, with follow-on investments from NVIDIA and other institutions. According to The Wall Street Journal, Mirendil’s valuation at that time reached $1 billion (approximately RMB 6.672 billion).
In terms of compute infrastructure, Mirendil signed a multi-year strategic partnership agreement with Google Cloud in August 2026.
According to disclosures from both parties, the collaboration value exceeds $100 million (approximately RMB 667 million), primarily covering Google TPU and NVIDIA GPU compute resources. This agreement provides stable infrastructure support for Mirendil’s large-scale model pre-training and agent experimentation.

▲ Mirendil’s official partnership announcement (Source: X)
October 21, 2026, Mirendil launched its first foundational model, Mirendil-Base, with parameter count ranging from approximately 100 billion to 500 billion.
Officially claimed, the model was pre-trained from scratch by one researcher with agent assistance, completed in only 3.5 months, with total compute cost below $10 million (approximately RMB 6.67175 million), inclusive of experiments, agent token usage, and data processing.
Agents completed over 99% of the coding tasks, handling infrastructure setup, algorithm selection, data collection, and iterative refinement; human researchers were primarily responsible for strategic direction and outcome evaluation.
Subsequently, the company conducted post-training based on Mirendil-Base, releasing Fern 0.1. This model has parameters ranging from 0.1 trillion to 0.5 trillion, supporting context windows up to approximately 10 million tokens.
The model is tailored for high-capacity tasks such as debugging training logs, analyzing full agent trajectories, and auditing pre-training data. It is currently in early access application phase.
Additionally, Mirendil announced ongoing development of larger parameter variants: Mirendil-Base-Large (approximately 500 billion to 2 trillion parameters) is about 80% complete, while Mirendil-Base-XL (approximately 2 trillion to 10 trillion parameters) is around 20% complete.

▲ Model progress announcement
Mirendil’s core product—the multi-agent R&D platform Lor—entered private beta testing on October 28, 2026. Mirendil revealed that internal R&D operations are now fully powered by Lor.
The platform can simultaneously orchestrate hundreds of agents to write code, run experiments, and automatically request and release compute resources, enabling continuous operation for days or even weeks.

▲ During one run, Lor’s agents built an environment from 100 PyTorch pull requests: 102 agents, averaging 8 minutes per agent

▲ Agents dynamically allocate compute resources and maintain full control over each virtual machine and node they launch

▲ During a nighttime run, Lor’s agents discovered a security vulnerability in LiteLLM and completed patching before dawn
In Lor, agents are decoupled from the physical devices they operate on. Agents determine what to run, and Lor sends each command to user-selected devices. Moreover, agents only operate on machines explicitly added by users, who can remove any device at any time.

▲ Agent-device separation
Currently, Lor supports macOS, Linux, and web platforms, with Windows, iOS, and Android versions planned for release. The platform is compatible with open-source models such as Kimi and DeepSeek, as well as cutting-edge proprietary models including Claude, GPT, Gemini, Grok, and Mirendil’s own models.

▲ Compatible models
Mirendil was co-founded by four individuals whose core team members come from Anthropic, Google DeepMind, xAI, and OpenAI, bringing direct experience in model inference, automated R&D, engineering execution, and early-stage research.
Co-founder and CEO Behnam Neyshabur holds a bachelor’s and master’s degree from Sharif University of Technology, and earned his Ph.D. in Computer Science from Toyota Technological Institute at Chicago. He then pursued postdoctoral research under Yann LeCun at New York University and Princeton Institute for Advanced Study.
From 2019 to 2024, Behnam served as a Senior Research Scientist at Google DeepMind, co-leading the Blueshift team focused on enhancing foundation model reasoning capabilities. The team directly contributed to the development of Minerva, the first language model demonstrating strong performance in mathematical reasoning, and was responsible for early inference capability design in the Gemini series. His later work also informed improvements in Gemini 1.5 and Gemma 2.
From late 2024 to December 2025, Behnam co-led the Discovery team at Anthropic, aiming to build “AI scientists/engineers” capable of long-term reasoning and scientific tasks.
In academia, he is a co-author of the ICLR 2021 paper “Sharpness-Aware Minimization” (SAM optimizer), a technique widely used to enhance generalization in deep learning models. His research has been cited over 50,000 times on Google Scholar.

▲ Behnam Neyshabur (Source: GitHub)
Co-founder and CTO Harsh Mehta is a colleague of Behnam from Google to Anthropic. He graduated from IIT Guwahati and spent nearly a decade at Google Search and Google DeepMind, contributing to the development of the Gemini series models, with publications in 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 initial version of the company’s internal automated AI research platform (autoresearch). This platform aimed to accelerate pre-training and experiment iteration using models themselves—an immediate precursor to Mirendil’s current technical direction.

▲ Mirendil founding team photo, Harsh Mehta on the far right (Source: wsj)
Co-founder Shayan Salehian earned his undergraduate degree from Sharif University of Technology and a master’s from the University of California, Riverside.
From 2018 to 2024, he worked as a Machine Learning Engineer at Twitter and X, contributing to timeline recommendation systems, model serving, and large-scale system engineering.
From October 2024 to early 2026, Shayan joined xAI, working on post-training, inference, memory systems, coding infrastructure, and agent-related components for the Grok series 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 in the International Olympiad on Astronomy and Astrophysics (IOAA), winning a silver medal. She then entered MIT to study Electrical Engineering and Computer Science. While at MIT, she served as a student researcher at OpenAI, focusing on model interpretability and program synthesis.

▲ Tara Rezaei (Source: Tara’s personal website)
Today, frontier AI capabilities remain concentrated in a small number of labs with massive compute budgets, elite talent, and full experimental feedback loops. Mirendil seeks to disrupt this structure using agents, compressing what once required dozens of people, months, or even years of effort into a single person, a few months, and a sub-$10 million compute budget.
Specialized labs in drug discovery, materials science, robotics, and similar fields may now conduct near-frontier model iterations without assembling a full pre-training team—potentially giving rise to a new generation of “domain-specific mini-labs.” Capital has invested twice within half a year, likely betting on precisely this potential.
However, Mirendil’s most impressive achievements so far remain internal experiments. How will external users experience it? And in the era of rapid iteration among leading large models, how will the startup establish sustainable technical moats? These are critical questions facing the founding team.
Source: SmartThings
Disclaimer: Contains third-party opinions, does not constitute financial advice
The AlphaGo Moment in Cybersecurity, or Starting from These 466 Agent Matchups
12 days ago
Yushu Surges 629% on First Day of Trading, Humanoid Robots Sell 5,500 Units
08-19
Yushu's 61 Billion IPO: Who Will Make Money?
08-07
AI stocks down 40%-60%, yet NVIDIA's valuation is at a 10-year low
08-06
HBM Bottleneck Migration, NVIDIA Also Aims to Downgrade
08-06
AI Servers Fuel Demand for Minor Metals: Prices Surge Over 70% in Half a Year
08-06
AI is not a search bar—it's a supercomputer costing just a few dozen dollars
08-05






