Mirendil

Mirendil is an artificial intelligence research company based in San Francisco that trains large language models specialized for AI research and engineering work, and builds systems around them that are meant to automate more and more of an AI lab's research loop.^[1][10][22] It was founded in 2026 by Behnam Neyshabur, its chief executive, and Harsh Mehta, its chief technology officer, both former researchers at Anthropic, together with co-founders Shayan Salehian, formerly an engineer at xAI, and Tara Rezaei, a former student researcher at OpenAI.^{[4][6][13][18][23]} The company calls its approach "self-accelerating AI R&D" and says it wants to make frontier AI research and development available to scientists and engineers outside the largest labs, in fields such as drug discovery, chemistry, biology and robotics.^[1]

Mirendil announced a $200 million seed round in June 2026, led by Andreessen Horowitz and Kleiner Perkins and followed by an investment from Nvidia; TechCrunch and SiliconANGLE reported that the round valued the company at $1 billion.^[2][7][10] In August 2026 it signed a multi-year compute partnership with Google Cloud that Neyshabur told TechCrunch was worth more than $100 million.^[10][11] In September 2026 Bloomberg reported that the company was in talks to raise up to $1 billion at a $5 billion valuation.^[13] In October 2026 Mirendil described its first pretrained model, Mirendil-Base, which it said one researcher working with AI agents had built in 3.5 months for less than $10 million in compute, and announced Fern 0.1, a post-trained version of that model with a context window of up to 10 million tokens.^[14][15]

Overview

Name

Mirendil's website glosses the name as "friend of precious things, the hidden and undiscovered treasures, i.e. friend of discovery," and calls the company "the first lab from the future."^[1] Gizmodo described the name as Elvish, adding that it joined a growing list of technology company names inspired by The Lord of the Rings, and The Next Web called it "a name from Tolkien."^[8][9]

Mission and approach

The company's founding statement says that "for the first time, there are real signs of AI being able to do meaningful AI research," and that the lab which automates this loop "will build the thing that is capable of continually improving itself toward any goal."^[1] Mirendil describes itself as "a frontier lab building systems that excel at AI R&D": it trains models that are strong at AI research and engineering, then redesigns the rest of the lab around them to make the loop "faster, more capable, and more autonomous."^[1] Its stated reason is access. In the company's words, a lab that wants to use AI in drug discovery, chemistry, biology or robotics today "must also become a frontier AI lab," which is expensive and depends on expertise concentrated in a few organizations; Mirendil says it wants to make that capability widely available instead.^[1] In August 2026 it described the long-term aim as letting teams in medicine, biology and materials science build specialized AI systems "without needing the infrastructure of the world's largest AI labs."^[11]

Andreessen Horowitz partners Matt Bornstein and Malika Aubakirova, announcing their firm's investment, described the intended product as "a lab-grade research platform to help normal engineers do frontier AI work."^[3] They wrote that the system is meant to "loop over research and engineering problems more autonomously," that engineers and AI researchers would probably be its first users, and that the goal is to support less technical users so that scientists and other domain experts can run their own experiments. They called the idea "vibe research."^[3] The post argued that most large AI labs had built some version of such a platform for internal use but had little economic incentive to release it to competitors, so an independent company should take on the problem.^[3] In his launch announcement Neyshabur wrote that "the most important application of AI is AI itself," and argued that spreading the capability beyond a few labs would let businesses and science labs "own their AI and infrastructure."^[23]

Press coverage has tied the effort to recursive self-improvement, the idea of AI systems that iteratively improve themselves.^[10] Neyshabur told TechCrunch, "You can have a self-improving AI where you can point a problem at it and it keeps getting better with time," and gave research on Alzheimer's disease as an example of the kind of ambitious goal such a system could keep working on.^[10] The Next Web reported that the founders consider the approach the "shortest path" to faster science and a problem that can be supervised rather than avoided.^[8] In October 2026 the company wrote that its next model "will be built partly by Mirendil models," while adding: "We don't think we have reached fully autonomous recursive self-improvement yet, but we are closer than we have ever been."^[14] Mehta wrote at the same time that Mirendil planned to release the tools it had used to pretrain its first model "so more people can do it too."^[25]

Early reporting described the company in more applied terms. The Information reported in March 2026 that the startup aimed "to conduct AI-driven research and development for scientific fields such as biology and materials science."^[5] SiliconANGLE reported in June 2026 that Mirendil's job postings described work on new variants of the transformer architecture, including novel attention mechanisms, and on reinforcement learning sandboxes.^[7]

History

Founding

Neyshabur and Mehta met at Google in 2019; Kleiner Perkins wrote that their first contact was a cold email, and that the two later worked together on pretraining and on reasoning in math and code for Gemini.^[4][8] Both joined Anthropic in late 2024 and left in December 2025.^[6][8] Neyshabur's website dates his role as Mirendil's co-founder and chief executive from 2026, and Bloomberg described the company as started in 2026 by former Anthropic staff.^[13][18]

On March 13, 2026, The Information reported that former Anthropic researchers were in talks to raise $175 million at a $1 billion valuation for a new startup.^[5] The Decoder, reporting the next day and citing The Information for the funding talks, identified the company as Mirendil, named Neyshabur as chief executive and Mehta as chief technology officer, and said Andreessen Horowitz and Kleiner Perkins were reportedly co-leading the round, with terms not yet final.^[6] The Decoder placed Mirendil in a wave of so-called neo-labs, specialized AI startups founded by researchers who had left large AI companies.^[6]

Launch and seed round

Mirendil formally announced itself on June 24, 2026.^[2] It said its founding team consisted of 20 researchers and engineers from institutions including Anthropic, xAI, Google DeepMind and OpenAI, and that Andreessen Horowitz and Kleiner Perkins had led a $200 million seed round, "followed by an investment from NVIDIA among others."^[2] Both lead investors published announcements the same day, Andreessen Horowitz's written by Bornstein and Aubakirova and Kleiner Perkins's by Mamoon Hamid.^[3][4] The company's announcement did not state a valuation; TechCrunch and SiliconANGLE reported that the round valued Mirendil at $1 billion.^[7][10] The Next Web called it one of the largest seed rounds the sector had seen and noted that the company had no product at the time.^[8] Mehta, in his own announcement, described Mirendil as "a small team in SF."^[23]

Google Cloud partnership

On August 6, 2026, Mirendil announced a multi-year partnership with Google Cloud to run its training, inference and AI research workloads on Google Cloud's AI Hypercomputer, using both Google's Tensor Processing Units and Nvidia GPU infrastructure.^[11] TechCrunch, which reported the deal as an exclusive, said Neyshabur put its value at more than $100 million, roughly half of what the company had raised in its seed round.^[10] Google said it had worked with Mirendil on the design and deployment of combined TPU and Nvidia infrastructure across compute, storage, networking and control planes, that the setup used managed training clusters running in Gemini Enterprise Agent Platform, and that Mirendil was "already live with a cluster of TPU v5P chips," with Nvidia systems coming online soon.^[12] Mehta told TechCrunch that access to several kinds of chips lets the company "mix and match workloads with the right kind of accelerators."^[10]

Funding talks

On September 22, 2026, Bloomberg reported, citing people familiar with the matter, that Mirendil was in talks to raise new funding at a $5 billion valuation.^[13] According to PYMNTS's account of the report, Kleiner Perkins was in talks to lead a round that could bring in up to $1 billion, Andreessen Horowitz was also in discussions to invest, the company had more than 20 staff, and the new capital would pay for Mirendil's own model and future training runs, with a frontier model for engineering and research work planned by early 2027.^[13] Mirendil's news page listed no announcement of a new round as of October 27, 2026.^[24]

First models

On October 21, 2026, Mirendil described Mirendil-Base, its first pretrained model, and on October 26 it announced Fern 0.1, the first post-trained version of that model.^[14][15][16]

Models

Mirendil-Base

Mirendil presented Mirendil-Base as the result of an experiment testing whether "one researcher, working with agents," could pretrain a competitive coding model.^[14] According to the company, one researcher built it in 3.5 months with less than $10 million in compute, a figure it says includes all research and small-scale experiments, the final training run, the agents' token spending and data processing.^[14] Mirendil said AI agents wrote more than 99% of the code, ran experiments and diagnosed failures, while the researcher "decided what to try and what to trust, and intervened when needed."^[14] The project started from an empty repository, and the training infrastructure, architecture, training algorithm and data pipelines were written from scratch on top of general libraries such as PyTorch and JAX.^[14] The company said all training data came from GitHub and other public code, web pages and public datasets, with no purchased or licensed data, no distillation and no synthetic data.^[14] It credited efficiency gains to an improved architecture, tuning of the data mixture and kernel optimization.^[14] These figures are the company's own; the post gives parameter ranges rather than exact counts and does not describe the architecture or the number of training tokens.^[14]

Mirendil-Base is the first and smallest of three coding-specialized base models, designed to do well on the coding work done inside a frontier AI lab.^[14][16]

The company measured cross-entropy in bits per byte, a loss per byte of text that lets models with different tokenizers be compared, on held-out code.^[14] On code from Mirendil's own repositories it reported the following, noting that Kimi K3, GLM-5.3 and Inkling are post-trained models, which generally score slightly worse on this measure than base models:^[14]

Mirendil listed all five comparison models as mixture-of-experts models with between 550 billion and 2.8 trillion total parameters.^[14] On seven external repositories published after the models were trained, Mirendil-Base averaged 0.222 bits per byte, behind Kimi K3 (0.193) and ahead of GLM-5.3 (0.242), DeepSeek-V4-Pro-Base (0.243) and Nemotron-550B-Base (0.247).^[14] Jack Clark's Import AI newsletter summarized the result as Mirendil-Base outperforming Nemotron-550B-Base, GLM-5.3 and DeepSeek-V4-Pro-Base while underperforming Kimi K3, with both larger models in training already ahead of K3.^[17]

Fern

On October 26, 2026, Mirendil announced Fern as its first model, pretrained from scratch with a context window of up to 10 million tokens.^[15][16] Its first version, Fern 0.1, is "Mirendil-Base after a short round of post-training."^[15] The company says the window holds about 23 MB of code, roughly half a million lines (a codebase on the scale of the Megatron-LM library), dozens to hundreds of complete agent traces, or days of an agent's work.^[15] It positions Fern for "focused, high-volume agentic tasks" in AI development: auditing a pretraining shard for duplicates, benchmark contamination and junk; debugging a training run from its logs; reading agent or training traces end to end; and agentic coding on research code such as training loops, data pipelines and evaluations.^[15] Mehta wrote that the company had "a lot more work to do on making it as usable as 1M context."^[26]

Mirendil published early agentic-coding results for Fern 0.1, which it ran itself with the OpenCode harness and one attempt per task.^[15] Scores for the other models came from their model cards where available, then from benchmark or independent leaderboards, often with other harnesses and effort settings; where none existed, Mirendil ran the evaluation itself.^[15] Selected results (pass rate, %):

The same chart also included Inkling and Nemotron 3 Ultra, which scored below Fern 0.1 on DeepSWE, Terminal-Bench and SWE Atlas.^[15]

For long context, Mirendil measured next-token prediction on its own repositories, concatenated into one 42.9 MB stream and read with sliding windows.^[15] Fern's loss kept falling as the window grew, from 0.2456 bits per byte at a 32 KiB window to 0.1552 at a 23 MiB window (about 10.2 million tokens); the other models were evaluated up to about 1 million tokens.^[15] In the same chart, Kimi K3's loss at its longest evaluated window (about 0.9 million tokens) was 0.1395, lower than Fern's at 10.2 million tokens.^[15]

The company also ran five tasks from the RULER long-context benchmark on Mirendil-Base, the pretrained model behind Fern 0.1, with greedy decoding, thinking turned off and exact string matching. Each cell averages 100 examples, except 20 at 10 million tokens:^[15]

Founders and team

Much of Neyshabur's earlier research concerned why deep neural networks generalize.^[4][18] He is a co-author of Sharpness-Aware Minimization (SAM), a 2020 training procedure that seeks parameters lying in neighborhoods of uniformly low loss in order to improve generalization,^[19] and of the 2022 paper that introduced Minerva, a language model further trained on technical content to solve quantitative reasoning problems.^[20] His website says the Blueshift team he co-led introduced Minerva in 2022, was responsible for the reasoning capabilities of the first version of Gemini in 2023, and developed a math-specialized Gemini 1.5 model in 2024; it also says his Anthropic team's work led to a threefold improvement in the computer-use capability of Claude Sonnet 4.5.^[18] At Google Research, Mehta and Neyshabur co-authored several papers,^[18] including one introducing the 2022 Gated State Space layer, a state space model layer for autoregressive modeling of books, GitHub code and arXiv mathematics that showed zero-shot generalization to longer inputs.^[21]

Neyshabur's website lists Preetum Nakkiran, Xiaoxia Wu and Anna Golubeva, researchers he previously mentored, as members of technical staff at Mirendil.^[18]

Funding

Reception

Kleiner Perkins's Hamid wrote that the founders were "building the system that builds systems."^[4] Jack Clark, a co-founder of Anthropic, wrote in Import AI that Mirendil-Base was "less a technical breakthrough and more a proof-of-concept for how AI development could work in the future," showing that a single person working with AI could build systems that a year earlier would have taken teams of ten or more.^[17] He added that the post demonstrated how a new organization could, "from a cold start, get close to the open weight frontier" for a model of its size.^[17]

Other coverage stressed how early the company was. The Next Web summarized the pitch as selling "the self-improving AI that the big labs build for themselves and guard from everyone else," and wrote that whether the democratization vision "survives contact with a real product is another matter."^[8] Gizmodo noted that recursive self-improvement is regarded by some as a path to runaway AI, and reported Mirendil's view that the main problem is that access to such capabilities is gated by a few well-funded labs.^[9] PYMNTS observed that the reported $5 billion valuation would be 25 times the company's funding to date, at a time when investors were pricing unproven AI labs "on research ambition rather than revenue."^[13] As of October 27, 2026, Mirendil's model results were self-reported and Fern was not yet generally available.^[15][16]

See also

• List of neolabs
• Neolab
• Recursive self-improvement
• Recursive Superintelligence
• Periodic Labs
• Thinking Machines Lab
• AI for Science

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Mirendil, "Mirendil", company website, accessed October 27, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6Mirendil, "Announcing Mirendil", June 24, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Matt Bornstein and Malika Aubakirova, "Investing in Mirendil", Andreessen Horowitz, June 24, 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7Mamoon Hamid, "Mirendil: Building the system that builds systems", Kleiner Perkins, June 24, 2026.
5. ^1 ^2 ^3Julia Hornstein, Anissa Gardizy and Sri Muppidi, "Ex-Anthropic Researchers Are Raising Capital For New Startup at $1 Billion Valuation", The Information, March 13, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Matthias Bastian, "Ex-Anthropic researchers launch AI startup Mirendil to tackle scientific research", The Decoder, March 14, 2026.
7. ^1 ^2 ^3 ^4 ^5 ^6Maria Deutscher, "Mirendil raises $200M to speed up scientific research with AI", SiliconANGLE, June 25, 2026.
8. ^1 ^2 ^3 ^4 ^5 ^6Cristian Dina, "Ex-Anthropic researchers raise $200M for self-improving AI", The Next Web, June 29, 2026.
9. ^1 ^2Webb Wright, "Don't Be Afraid of Self-Improving AI, Says a16z-Backed Startup Mirendil", Gizmodo, June 29, 2026.
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Rebecca Bellan, "Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI", TechCrunch, August 6, 2026.
11. ^1 ^2 ^3Mirendil, "Scaling self-accelerating AI with Google", August 6, 2026.
12. ^1 ^2Darren Mowry, "Mirendil taps AI Hypercomputer TPUs and GPUs for pre- and post-training applications", Google Cloud Blog, August 2026.
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9PYMNTS, "Anthropic Alumni Startup Sees Valuation Jump Fivefold", September 22, 2026 (reporting on a Bloomberg article of the same date).
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19Mirendil, "Pretraining at scale without a pretraining team", October 21, 2026.
15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25Mirendil, "Introducing Fern", October 26, 2026.
16. ^1 ^2 ^3 ^4 ^5Mirendil (@mirendil), "Introducing Fern, Mirendil's first model", post on X, October 26, 2026.
17. ^1 ^2 ^3Jack Clark, "Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means", Import AI, October 26, 2026.
18. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Behnam Neyshabur, "Behnam Neyshabur", personal website, accessed October 27, 2026.
19. ^Pierre Foret, Ariel Kleiner, Hossein Mobahi and Behnam Neyshabur, "Sharpness-Aware Minimization for Efficiently Improving Generalization", arXiv:2010.01412, October 3, 2020.
20. ^Aitor Lewkowycz et al., "Solving Quantitative Reasoning Problems with Language Models", arXiv:2206.14858, June 29, 2022.
21. ^1 ^2Harsh Mehta, Ankit Gupta, Ashok Cutkosky and Behnam Neyshabur, "Long Range Language Modeling via Gated State Spaces", arXiv:2206.13947, June 27, 2022.
22. ^1 ^2 ^3Mirendil, "Mirendil", company page, LinkedIn, accessed October 27, 2026.
23. ^1 ^2 ^3 ^4 ^5Harsh Mehta, "I'm excited to announce what we've been working on for the past few months: Mirendil", LinkedIn post resharing Behnam Neyshabur's launch announcement, June 24, 2026.
24. ^1 ^2Mirendil, "News", accessed October 27, 2026.
25. ^Harsh Mehta, "I've worked on pretraining LLMs for a long time", LinkedIn post, October 21, 2026.
26. ^Harsh Mehta, "More updates", LinkedIn post, October 26, 2026.