Mirendil releases a hundred-billion-parameter coding foundation model Mirendil-Base, trained by 1 researcher and an Agent

How many people does it take to train a 500-billion-parameter large model?

But a Silicon Valley startup called Mirendil has compressed that number to the extreme—just one person.

Just now, Mirendil published an article on its official website with a very sensational title: "Large-Scale Pretraining Without a Pretraining Team."

Below the article title are three numbers:

1 researcher, three and a half months, cost under $10 million.

This money covered the Agent's token consumption, data processing, and all research experiment costs. With only these inputs, they really did train a hundred-billion-parameter model: Mirendil-Base, a coding-specialized base model in the 100-billion to 500-billion parameter range.

Mirendil stated in the article that more than 99% of the new model's code was not written by humans, but by Agents.

The human researcher only needs to be responsible for determining direction, making judgments, and stepping on the brakes at key moments. Tasks such as writing code, running experiments, and troubleshooting errors were all handed over to Agents.

In Mirendil's tests, some metrics of the new model have already approached mainstream frontier models such as Kimi K3, GLM-5.3, and DeepSeek-V4.

Is the era of "single-handedly building a base model" really about to arrive...?

01

According to the official account, Mirendil-Base is the first product in the Mirendil series of models, positioned as a coding-specialized base model, with target users being researchers and engineers inside frontier AI labs.

To verify whether one person can actually build a model, Mirendil set many harsh conditions. For example, they could not use off-the-shelf pretraining frameworks, nor could they inherit code.

The researcher and Agents started from scratch and built everything over more than 100 days—training infrastructure, model architecture, training algorithms, and data pipelines were all written by themselves, only referencing some popular open-source projects as templates on general-purpose libraries such as PyTorch and JAX.

Of course, "themselves" here usually refers to the Agents.

The researcher was mainly responsible for deciding what to try, what to trust, and intervening when necessary. Specific actions such as writing code, running experiments, and diagnosing errors were basically all run by the Agents themselves.

On the data side, Mirendil used entirely public data, such as GitHub open-source code, public web pages, and public datasets. The company emphasized that during training it did not buy a single line of data, did not perform distillation, and did not use synthetic data.

In addition, Mirendil hoped to improve efficiency through design improvements. "We achieved significant efficiency gains through an improved architecture, advanced tuning of data mixing, other data-level improvements, and kernel optimization," the announcement said.

The final product is Mirendil-Base, with 100 billion to 500 billion parameters.

So what level is a hundred-billion-parameter model built by one person actually at?

Mirendil compared five models: Kimi K3, GLM-5.3, Nemotron-550B-Base, DeepSeek-V4-Pro-Base, and Inkling, all large-parameter models with MoE architectures.

In the cross-entropy test on new code repositories (where the model guesses how the code is written; the lower the score, the more accurately it guesses), the comparison results of the models were as follows:

Kimi K3 ranked first with a score of 0.193; Mirendil-Base ranked second with a score of 0.222; followed by GLM-5.3, DeepSeek-V4, and Nemotron.

In other words, in some coding dimensions, a pretrained model trained by one person and Agents working together can already go toe-to-toe with finished models from leading major companies.

However, it must be pointed out that, under normal circumstances, base models perform better than post-trained models in cross-entropy tests. As a base model, Mirendil-Base's score is actually higher than the post-trained Kimi-K3, which shows that Kimi-K3's true strength is much stronger than its surface score suggests, and the gap between Mirendil-Base and it is larger than it appears.

According to the official introduction, above Mirendil-Base there are two more tiers: Large at 500 billion to 2 trillion, and XL directly at 2 trillion to 10 trillion.

At present, Large has been trained 80%, and XL has just begun, at only 19%, yet it is already close to catching up with Large's current score.

In other words, with the backing of money and Agents, one person training a trillion-parameter model is not a dream—a threshold that many major AI companies have not yet reached.

02

Pretraining hundred-billion and even trillion-parameter large models is extremely expensive, time-consuming, and requires a large number of people.

Among hundreds of researchers, some study what the model should look like and how to do scaling; some manage the cleaning and mixing of trillions of tokens; and some watch over thousands, or even hundreds of thousands, of compute cards to ensure training does not stop.

For example, Anthropic has already gone far down this path.

It previously disclosed that in May of this year, more than 80% of the merged code in its codebase was written by Claude.

Scientists, researchers, and small teams who want to train a model themselves previously found it almost impossible.

But Mirendil-Base has proven that, under certain conditions, a single person can also "hand-build" a model.

In a post on its official website, Mirendil stated that until recently, the bottleneck in pretraining a competitive model was still people: an expert team that took years to assemble.

Now, that situation is changing.

What once only large research institutions could accomplish can now be done by a single researcher. When Agents take on most of the hands-on work, the bottleneck shifts to compute, data, and the judgment of the person directing the Agent.

"Anyone can train a model" carries enormous imaginative potential.

Previously, most people and companies that needed a dedicated model would often rent a general-purpose model and then adapt it at the edges.

Now, a chip design team, a robotics lab, or a hospital can train a model with its own data for its own problems. Base models and the systems built around them will no longer be something only a few big companies can produce; they will become far more widespread.

On the other hand, AI projects that previously required a large team to launch can now start with a small team, or even one person; and as compute becomes cheaper and more powerful, more people will be able to try these projects. "One-person companies" or "super-individuals" are expected to become a reality.

People's way of working will also change: as more and more work is handed over to Agents, humans' daily work will look less and less like "doing things by hand" and more and more like "leading a small laboratory": setting research directions, designing experiments worth running, and judging which ideas are worth pursuing.

Mirendil paints an even more dazzling picture: the next model will be built in part by Mirendil's own models, and every model that becomes stronger at building AI will make building the next model cheaper and faster.

"We believe we have not yet reached fully autonomous RSI, but we are closer than ever before," the company said.

In addition, the company has already begun post-training Mirendil-Base. Mirendil-Base-Large and the frontier-level Mirendil-Base-XL are still in training.

03

Behind Mirendil are two former Anthropic employees: Behnam Neyshabur and Harsh Mehta.

Neyshabur was born in the 1980s, was born in Iran, and attended university there. In 2011 he went to the United States and earned a PhD at the Toyota Technological Institute at Chicago. After graduation, he worked at Google DeepMind for five years, joined Anthropic at the end of 2024, and led the Discovery team, specifically researching how to get AI to participate in R&D like a scientist.

Mehta is of Indian descent, graduated from IIT Guwahati, and was Neyshabur's colleague at DeepMind. After joining Anthropic, Mehta built the first version of Anthropic's internal automated AI research platform.

In December 2025, the two left their jobs one after the other and founded Mirendil, whose positioning is to let AI help scientists build their own AI, rather than merely using AI to assist science.

In June of this year, Mirendil officially announced a $200 million seed round at a post-money valuation of $1 billion. Investors include a16z, Kleiner Perkins, Nvidia, and others.

The team currently has about 20 people, all researchers and engineers from Anthropic, xAI, Google DeepMind, and OpenAI, working in downtown San Francisco.

The debut of Mirendil-Base marks the company's crossing of its first milestone.

But then again, "a single person training a hundred-billion-parameter model" sounds wonderful, but the barrier has not been lowered by much.

A cost of $10 million is nothing to the big companies, and Mirendil, which has just raised funding, can still afford it. But for the vast majority of individuals and small teams, it is still an astronomical figure.

Not to mention that being able to direct Agents to pull off such a big job is not something just any engineer can pick up. Mirendil's talent quality and wealth of AI project experience are in the first tier across all of Silicon Valley.

The era of "one person hand-building a large model" is indeed approaching, but it is still far from ordinary people.

The release of Mirendil-Base is the first signpost on this long road. As for whether it can truly lower the barrier to large models, that will depend on the actual level of subsequent larger-scale models and whether this methodology can be replicated by others.