For years, building a frontier-scale artificial intelligence model has been associated with enormous research teams, specialized engineering departments, and hundreds of millions of dollars in infrastructure.

Mirendil is challenging part of that assumption.

On October 21, 2026, the company described how a single researcher, directing AI coding agents, trained a large language model from scratch in approximately three and a half months.

The model, called Mirendil-Base, reportedly contains between 100 billion and 500 billion parameters. According to the company, AI agents wrote more than 99% of the code required for its development.

The reported training cost was under $10 million.

These are remarkable claims, but the most interesting part is not the parameter count or even the cost.

It is the possibility that the organizational structure required to train foundation models is changing.

If AI agents can perform most of the implementation work, the next generation of AI laboratories might not need enormous engineering departments to develop their own models.

They may instead need a small number of exceptionally capable researchers, reliable agent infrastructure, access to high-quality data, and significant computing resources.

That would change who gets to participate in the foundation model industry.

Why Training Foundation Models Has Traditionally Required Large Teams

Training a large language model is fundamentally different from building an application that calls an existing AI API.

A typical AI application relies on a model that has already been trained. Developers build interfaces, retrieval systems, tools, and workflows around that model.

Pretraining requires creating the model itself.

This involves several demanding engineering disciplines:

• Collecting, filtering, and deduplicating training data.
• Designing model architectures and tokenization strategies.
• Implementing distributed training across accelerators.
• Managing memory utilization and communication overhead.
• Detecting numerical instability and failed training runs.
• Evaluating model quality throughout training.
• Managing checkpoints, storage, and recovery.
• Optimizing expensive hardware utilization.

Each of these activities can involve substantial engineering work.

In traditional AI laboratories, different teams often specialize in different parts of this pipeline.

Infrastructure engineers optimize distributed systems. Researchers investigate model architectures and training techniques. Data engineers manage enormous datasets. Evaluation specialists measure whether the model is improving.

Coordinating these activities becomes increasingly difficult as the training operation grows.

This creates two distinct costs: the computational cost of training and the organizational cost of building and operating the training system.

Mirendil's announcement suggests that AI agents may dramatically reduce the second.

What Mirendil Claims to Have Achieved

According to Mirendil's October 21 announcement, the company developed Mirendil-Base using publicly available training data, without relying on outputs copied from existing models.

The company describes the project as a from-scratch pretraining effort.

Its headline claims are substantial:

Metric	Reported result
Model	Mirendil-Base
Parameter count	Between 100B and 500B
Development time	Approximately 3.5 months
Human research leadership	One researcher
Code written by AI agents	More than 99%
Reported training cost	Under $10 million
Training data	Public data
Reported evaluation	Held-out code prediction

Mirendil says its model performs better than Nemotron-550B-Base and DeepSeek-V4-Pro-Base on the company's held-out code prediction evaluation.

It also reports that Mirendil-Base trails Kimi-K3 on that metric.

These results are interesting, but they require careful interpretation.

The comparison is based on a particular code prediction measurement selected and reported by Mirendil itself.

It does not establish that Mirendil-Base is superior to these models at general reasoning, software engineering, instruction following, or agentic coding.

No comprehensive task benchmark results were provided in the information available for this article.

Nevertheless, the reported achievement is significant enough to raise questions about how AI models could be developed in the future.

The Real Breakthrough May Be Organizational, Not Algorithmic

The AI industry typically measures progress through model performance.

A new model is considered important when it achieves better benchmark scores, supports larger context windows, or demonstrates improved reasoning capabilities.

Mirendil's announcement points toward another possible dimension of progress: the amount of human coordination required to build a model.

Consider a conventional pretraining organization.

Researchers propose experiments. Engineers implement training infrastructure. Other teams manage datasets, evaluate checkpoints, investigate failures, and improve performance.

Each activity requires communication and coordination.

Now imagine an alternative structure.

One researcher establishes the objectives, makes technical decisions, and evaluates results. AI agents implement code, write tests, investigate errors, and help maintain the training infrastructure.

The human researcher becomes less of an individual contributor writing every component and more of a technical director coordinating automated contributors.

The agents do not eliminate the need for expertise.

They change how expertise is applied.

Instead of spending most of the day implementing systems, the researcher can spend more time deciding which systems should exist, which experiments deserve resources, and which results are trustworthy.

That is potentially a much larger shift than simply making programmers more productive.

From Engineering Teams to Agent-Directed Research

A useful way to understand this transition is to compare two hypothetical workflows.

Traditional Pretraining Workflow

Research objective → Research team → Infrastructure team → Data engineering → Distributed training → Evaluation team → Research iteration

The work moves between specialized groups.

Each handoff introduces coordination overhead, and each experiment may require contributions from multiple people.

Agent-Directed Pretraining Workflow

Researcher → AI agents → Data and training infrastructure → Automated evaluation → Researcher review → Next iteration

In the second model, agents can perform implementation work in parallel while the researcher remains responsible for important decisions.

This is a conceptual illustration of an agent-directed workflow, not a verified description of Mirendil's internal architecture.

The distinction matters because writing code is only one component of research.

Agents may be able to implement a training system, but someone still needs to determine whether the experiment is scientifically meaningful.

Someone must decide whether an apparent improvement is real, whether the evaluation is valid, and whether another expensive training run is justified.

The bottleneck shifts from implementation capacity toward decision quality.

The Economics of Building Foundation Models Could Change

One of the most important implications concerns the economics of AI laboratories.

Historically, building a serious foundation model has required both substantial computational resources and an experienced engineering organization.

AI agents could reduce the cost of maintaining that organization.

However, they do not make computation free.

Training large models still requires accelerators, high-bandwidth networking, storage, electricity, and access to suitable training data.

These resources can remain extremely expensive even when software development becomes easier.

A simplified model of pretraining cost is:

Total cost = Compute + Data + Human labor + Infrastructure + Experimentation

AI agents primarily target human implementation labor and some operational overhead.

They may also reduce experimentation costs by accelerating debugging, automation, and infrastructure improvements.

But there is another possibility.

When experiments become easier to implement, researchers may choose to run more experiments.

The resulting increase in computational demand could offset some of the savings from automation.

In other words, cheaper engineering does not automatically mean cheaper research.

It may mean that more research becomes economically possible.

What Happens When Labs Buy Compute Instead of Hiring?

Suppose an AI startup previously needed dozens of engineers to create and maintain its training infrastructure.

If agents could reliably perform much of that implementation work, the startup might choose to employ fewer engineers and allocate more capital toward accelerators and datasets.

This would change the composition of its spending.

Rather than scaling primarily through headcount, it could scale through compute capacity.

Such a transition could create several effects.

First, smaller research organizations could attempt projects that previously required larger teams.

Second, demand for cloud GPUs and specialized AI infrastructure could increase.

Third, the value of experienced researchers capable of directing complex automated workflows could rise.

Fourth, companies might differentiate themselves through data quality, training methodology, and evaluation rather than the size of their engineering teams.

However, these outcomes are possibilities rather than established consequences of Mirendil's announcement.

A successful training run by one organization does not demonstrate that every laboratory can adopt the same structure.

Different model architectures, research objectives, hardware environments, and regulatory requirements create different organizational needs.

Why the Under-$10-Million Claim Needs Context

The reported cost is one of the most attention-grabbing parts of Mirendil's announcement.

Training a model containing hundreds of billions of parameters for less than $10 million would be notable.

But cost comparisons require consistent accounting.

Does the reported figure include hardware depreciation, failed experiments, dataset preparation, evaluation, storage, networking, and the cost of operating the AI agents?

Was it the cost of the final successful training run, or the entire research effort?

Were there cloud discounts, special hardware arrangements, or other financial advantages?

Without a detailed accounting methodology, it is difficult to compare the figure directly with costs reported by other AI laboratories.

There is also a distinction between the cost of producing a base model and the cost of producing a commercially useful assistant.

A pretrained base model may require additional post-training, instruction tuning, safety evaluation, tool integration, and deployment infrastructure.

Those activities introduce further expenses.

The under-$10-million figure is therefore best understood as Mirendil's reported pretraining cost, not a demonstrated all-inclusive price for building and operating a production-ready AI platform.

A Strong Code Prediction Score Is Not the Same as a Strong Coding Assistant

Mirendil's reported evaluation focuses on predicting held-out code.

This can provide useful information about how well a model learns programming language patterns and generalizes to unseen code.

However, software engineering involves much more than predicting the next token.

A practical coding assistant must understand requirements, navigate repositories, identify defects, use tools, write tests, and maintain consistency across long-running tasks.

These capabilities are not established by code prediction performance alone.

A model might perform well on held-out code while struggling with complex debugging or multi-file changes.

Similarly, a model optimized for general instruction following may behave differently from a base model evaluated primarily on language modeling objectives.

Independent benchmarks would help establish where Mirendil-Base actually stands.

Useful evaluations could include:

• Standardized code generation benchmarks such as HumanEval.
• Repository-level software engineering tasks such as SWE-bench.
• Long-context code understanding.
• General reasoning and mathematical problem-solving.
• Tool-use and agentic coding evaluations.
• Training reproducibility and independent testing.

Until broader results are available, the strongest supported conclusion is that Mirendil reports competitive performance on its chosen code prediction metric.

That is a narrower claim than saying it has built a frontier coding assistant.

The New Bottleneck Is Human Judgment

Perhaps the most important implication of agent-directed model development is the increasing value of technical judgment.

AI agents can generate code quickly.

But generating code and making the correct engineering decision are different problems.

In large-scale training, an incorrect configuration can waste enormous computational resources.

A subtle data processing error can contaminate an evaluation.

A distributed training bug can remain hidden until thousands of GPU hours have been consumed.

An apparent performance improvement may disappear when tested against a more representative dataset.

The person directing the agents must understand these risks.

They need enough knowledge to challenge outputs, interpret metrics, and recognize when automated systems are producing plausible but incorrect results.

This creates an interesting paradox.

As AI agents reduce the need for routine implementation work, deep technical expertise may become more important for the people supervising them.

The most valuable researcher may not be the person who writes the largest amount of code.

It may be the person who consistently makes the best decisions about what code should be written and how its results should be evaluated.

What This Means for Software Engineers

The implications extend beyond AI research laboratories.

Software development teams are already experimenting with coding agents that can implement features, generate tests, review changes, and investigate bugs, as described in reports from Anthropic and OpenAI.

Mirendil's reported workflow represents a more ambitious version of the same idea.

For a broader look at how cheaper software creation is reshaping where value lies, read AI Is Making Software Cheaper to Build. So What Is Actually Becoming More Valuable?.

Instead of using agents to accelerate individual programming tasks, the researcher reportedly used them to support an entire complex engineering effort.

For software engineers, this suggests that several capabilities could become increasingly valuable.

1. System Architecture

Understanding how components interact becomes more important when agents can implement individual components quickly.

2. Verification and Testing

Generated code still needs validation. Engineers who can design reliable tests and evaluations will remain essential.

3. Technical Decision-Making

Choosing the right tradeoffs between performance, reliability, cost, and complexity becomes a central skill.

4. Agent Orchestration

Developers may increasingly need to coordinate multiple automated workers, define responsibilities, and manage shared context.

For a practical example of building and orchestrating AI agents, see Lets look at Google Agent Development Kit (ADK).

5. Domain Expertise

Knowing what should be built, why it matters, and whether the result solves the intended problem becomes a differentiator.

The lesson is not that software engineers will disappear.

It is that the relationship between engineering effort and engineering output may change substantially.

Could This Democratize Foundation Model Development?

If Mirendil's approach proves reproducible, it could lower one of the barriers to entering the foundation model market.

Smaller research organizations might be able to train specialized models without building large engineering departments.

Universities, independent research groups, and well-funded startups could potentially experiment with architectures and training methods that previously appeared organizationally impractical.

But democratization has limits.

Access to high-performance compute remains uneven.

High-quality datasets can be difficult to obtain or prepare.

Evaluating large models requires technical expertise and additional resources.

And competing with the largest AI laboratories on general-purpose capabilities may still require enormous budgets.

Agent-directed development could lower the organizational barrier without eliminating the financial barrier.

It might produce a more diverse ecosystem of specialized models rather than making frontier-scale general-purpose training accessible to everyone.

That would still be a meaningful change.

The $200 Million Question

Mirendil has reportedly raised a $200 million seed round led by Andreessen Horowitz and Kleiner Perkins.

That financing introduces an interesting contrast.

The company claims to have completed its initial pretraining effort for under $10 million, yet it has raised substantially more capital.

The difference suggests a larger ambition, although the funding amount alone does not establish how the company intends to spend the money.

Potential investments could include additional compute, new training runs, improved datasets, post-training, evaluations, infrastructure, and recruitment.

The broader question is whether the next generation of AI laboratories will use capital differently.

Instead of employing enormous teams to build every part of the training stack, they may invest more heavily in compute and agent infrastructure while maintaining smaller research organizations.

If this approach succeeds, the competitive advantage of an AI laboratory could increasingly depend on how effectively a small number of researchers direct enormous amounts of automated work.

What Still Needs to Be Proven?

Mirendil's announcement is promising, but several important questions remain unanswered.

Can the reported results be reproduced independently?

How does Mirendil-Base perform on standardized coding and reasoning benchmarks?

What exactly is included in the reported training cost?

How much human intervention was required when agents encountered difficult engineering problems?

Would the same approach work for multimodal models or substantially larger training runs?

Can the model be turned into a reliable instruction-following assistant through post-training?

These questions matter because a compelling engineering demonstration is not automatically proof of a broadly applicable industry transformation.

The distinction between a successful individual experiment and a repeatable methodology is fundamental to scientific progress.

Final Thoughts

The most important idea behind Mirendil's announcement is not that one person can replace an entire AI research laboratory.

It is that AI agents may allow a single researcher to coordinate engineering work that previously required many specialists.

That possibility changes how we think about the economics of building foundation models.

Compute remains expensive. Data remains essential. Research judgment remains difficult to automate.

But if implementation becomes dramatically cheaper, the balance between human labor and computational resources could shift.

The next generation of AI laboratories may compete not only on model architectures and training budgets, but also on how effectively their researchers direct automated engineering systems.

Mirendil has not yet demonstrated broad model superiority through independent benchmarks.

What it has presented is a provocative claim about how a large model can be built.

And that may be the more consequential story.

The future of AI research might not be defined solely by how many researchers a laboratory employs.

It may also be defined by how much meaningful research each person can direct.

TL;DR

• Mirendil claims one researcher directed AI agents to train Mirendil-Base from scratch in approximately 3.5 months.
• The model reportedly contains between 100 billion and 500 billion parameters.
• AI agents reportedly wrote more than 99% of the project's code.
• The company reports a pretraining cost below $10 million and says it used public data without copying existing models' outputs.
• Mirendil reports outperforming Nemotron-550B-Base and DeepSeek-V4-Pro-Base on its held-out code prediction metric, while trailing Kimi-K3.
• These results have not established superiority on general coding tasks or independent comprehensive benchmarks.
• Agent-directed development could reduce engineering headcount requirements while increasing the importance of compute, data, and research judgment.
• The biggest potential shift is organizational: smaller teams may be able to coordinate increasingly complex AI research projects.

References

• Mirendil — Pretraining at Scale Without a Pretraining Team
• OpenAI — Evaluating large language models trained on code
• SWE-bench — Benchmarking software engineering agents

The One-Person Pretraining Team

For years, building a frontier-scale artificial intelligence model has been associated with enormous research teams, specialized engineering departments, and hundreds of millions of dollars in infrastructure.

Mirendil is challenging part of that assumption.

On October 21, 2026, the company described how a single researcher, directing AI coding agents, trained a large language model from scratch in approximately three and a half months.

The model, called Mirendil-Base, reportedly contains between 100 billion and 500 billion parameters. According to the company, AI agents wrote more than 99% of the code required for its development.

The reported training cost was under $10 million.

These are remarkable claims, but the most interesting part is not the parameter count or even the cost.

It is the possibility that the organizational structure required to train foundation models is changing.

If AI agents can perform most of the implementation work, the next generation of AI laboratories might not need enormous engineering departments to develop their own models.

They may instead need a small number of exceptionally capable researchers, reliable agent infrastructure, access to high-quality data, and significant computing resources.

That would change who gets to participate in the foundation model industry.

Why Training Foundation Models Has Traditionally Required Large Teams

Training a large language model is fundamentally different from building an application that calls an existing AI API.

A typical AI application relies on a model that has already been trained. Developers build interfaces, retrieval systems, tools, and workflows around that model.

Pretraining requires creating the model itself.

This involves several demanding engineering disciplines:

• Collecting, filtering, and deduplicating training data.
• Designing model architectures and tokenization strategies.
• Implementing distributed training across accelerators.
• Managing memory utilization and communication overhead.
• Detecting numerical instability and failed training runs.
• Evaluating model quality throughout training.
• Managing checkpoints, storage, and recovery.
• Optimizing expensive hardware utilization.

Each of these activities can involve substantial engineering work.

In traditional AI laboratories, different teams often specialize in different parts of this pipeline.

Infrastructure engineers optimize distributed systems. Researchers investigate model architectures and training techniques. Data engineers manage enormous datasets. Evaluation specialists measure whether the model is improving.

Coordinating these activities becomes increasingly difficult as the training operation grows.

This creates two distinct costs: the computational cost of training and the organizational cost of building and operating the training system.

Mirendil's announcement suggests that AI agents may dramatically reduce the second.

What Mirendil Claims to Have Achieved

According to Mirendil's October 21 announcement, the company developed Mirendil-Base using publicly available training data, without relying on outputs copied from existing models.

The company describes the project as a from-scratch pretraining effort.

Its headline claims are substantial:

Metric	Reported result
Model	Mirendil-Base
Parameter count	Between 100B and 500B
Development time	Approximately 3.5 months
Human research leadership	One researcher
Code written by AI agents	More than 99%
Reported training cost	Under $10 million
Training data	Public data
Reported evaluation	Held-out code prediction

Mirendil says its model performs better than Nemotron-550B-Base and DeepSeek-V4-Pro-Base on the company's held-out code prediction evaluation.

It also reports that Mirendil-Base trails Kimi-K3 on that metric.

These results are interesting, but they require careful interpretation.

The comparison is based on a particular code prediction measurement selected and reported by Mirendil itself.

It does not establish that Mirendil-Base is superior to these models at general reasoning, software engineering, instruction following, or agentic coding.

No comprehensive task benchmark results were provided in the information available for this article.

Nevertheless, the reported achievement is significant enough to raise questions about how AI models could be developed in the future.

The Real Breakthrough May Be Organizational, Not Algorithmic

The AI industry typically measures progress through model performance.

A new model is considered important when it achieves better benchmark scores, supports larger context windows, or demonstrates improved reasoning capabilities.

Mirendil's announcement points toward another possible dimension of progress: the amount of human coordination required to build a model.

Consider a conventional pretraining organization.

Researchers propose experiments. Engineers implement training infrastructure. Other teams manage datasets, evaluate checkpoints, investigate failures, and improve performance.

Each activity requires communication and coordination.

Now imagine an alternative structure.

One researcher establishes the objectives, makes technical decisions, and evaluates results. AI agents implement code, write tests, investigate errors, and help maintain the training infrastructure.

The human researcher becomes less of an individual contributor writing every component and more of a technical director coordinating automated contributors.

The agents do not eliminate the need for expertise.

They change how expertise is applied.

Instead of spending most of the day implementing systems, the researcher can spend more time deciding which systems should exist, which experiments deserve resources, and which results are trustworthy.

That is potentially a much larger shift than simply making programmers more productive.

From Engineering Teams to Agent-Directed Research

A useful way to understand this transition is to compare two hypothetical workflows.

Traditional Pretraining Workflow

Research objective → Research team → Infrastructure team → Data engineering → Distributed training → Evaluation team → Research iteration

The work moves between specialized groups.

Each handoff introduces coordination overhead, and each experiment may require contributions from multiple people.

Agent-Directed Pretraining Workflow

Researcher → AI agents → Data and training infrastructure → Automated evaluation → Researcher review → Next iteration

In the second model, agents can perform implementation work in parallel while the researcher remains responsible for important decisions.

This is a conceptual illustration of an agent-directed workflow, not a verified description of Mirendil's internal architecture.

The distinction matters because writing code is only one component of research.

Agents may be able to implement a training system, but someone still needs to determine whether the experiment is scientifically meaningful.

Someone must decide whether an apparent improvement is real, whether the evaluation is valid, and whether another expensive training run is justified.

The bottleneck shifts from implementation capacity toward decision quality.

The Economics of Building Foundation Models Could Change

One of the most important implications concerns the economics of AI laboratories.

Historically, building a serious foundation model has required both substantial computational resources and an experienced engineering organization.

AI agents could reduce the cost of maintaining that organization.

However, they do not make computation free.

Training large models still requires accelerators, high-bandwidth networking, storage, electricity, and access to suitable training data.

These resources can remain extremely expensive even when software development becomes easier.

A simplified model of pretraining cost is:

Total cost = Compute + Data + Human labor + Infrastructure + Experimentation

AI agents primarily target human implementation labor and some operational overhead.

They may also reduce experimentation costs by accelerating debugging, automation, and infrastructure improvements.

But there is another possibility.

When experiments become easier to implement, researchers may choose to run more experiments.

The resulting increase in computational demand could offset some of the savings from automation.

In other words, cheaper engineering does not automatically mean cheaper research.

It may mean that more research becomes economically possible.

What Happens When Labs Buy Compute Instead of Hiring?

Suppose an AI startup previously needed dozens of engineers to create and maintain its training infrastructure.

If agents could reliably perform much of that implementation work, the startup might choose to employ fewer engineers and allocate more capital toward accelerators and datasets.

This would change the composition of its spending.

Rather than scaling primarily through headcount, it could scale through compute capacity.

Such a transition could create several effects.

First, smaller research organizations could attempt projects that previously required larger teams.

Second, demand for cloud GPUs and specialized AI infrastructure could increase.

Third, the value of experienced researchers capable of directing complex automated workflows could rise.

Fourth, companies might differentiate themselves through data quality, training methodology, and evaluation rather than the size of their engineering teams.

However, these outcomes are possibilities rather than established consequences of Mirendil's announcement.

A successful training run by one organization does not demonstrate that every laboratory can adopt the same structure.

Different model architectures, research objectives, hardware environments, and regulatory requirements create different organizational needs.

Why the Under-$10-Million Claim Needs Context

The reported cost is one of the most attention-grabbing parts of Mirendil's announcement.

Training a model containing hundreds of billions of parameters for less than $10 million would be notable.

But cost comparisons require consistent accounting.

Does the reported figure include hardware depreciation, failed experiments, dataset preparation, evaluation, storage, networking, and the cost of operating the AI agents?

Was it the cost of the final successful training run, or the entire research effort?

Were there cloud discounts, special hardware arrangements, or other financial advantages?

Without a detailed accounting methodology, it is difficult to compare the figure directly with costs reported by other AI laboratories.

There is also a distinction between the cost of producing a base model and the cost of producing a commercially useful assistant.

A pretrained base model may require additional post-training, instruction tuning, safety evaluation, tool integration, and deployment infrastructure.

Those activities introduce further expenses.

The under-$10-million figure is therefore best understood as Mirendil's reported pretraining cost, not a demonstrated all-inclusive price for building and operating a production-ready AI platform.

A Strong Code Prediction Score Is Not the Same as a Strong Coding Assistant

Mirendil's reported evaluation focuses on predicting held-out code.

This can provide useful information about how well a model learns programming language patterns and generalizes to unseen code.

However, software engineering involves much more than predicting the next token.

A practical coding assistant must understand requirements, navigate repositories, identify defects, use tools, write tests, and maintain consistency across long-running tasks.

These capabilities are not established by code prediction performance alone.

A model might perform well on held-out code while struggling with complex debugging or multi-file changes.

Similarly, a model optimized for general instruction following may behave differently from a base model evaluated primarily on language modeling objectives.

Independent benchmarks would help establish where Mirendil-Base actually stands.

Useful evaluations could include:

• Standardized code generation benchmarks such as HumanEval.
• Repository-level software engineering tasks such as SWE-bench.
• Long-context code understanding.
• General reasoning and mathematical problem-solving.
• Tool-use and agentic coding evaluations.
• Training reproducibility and independent testing.

Until broader results are available, the strongest supported conclusion is that Mirendil reports competitive performance on its chosen code prediction metric.

That is a narrower claim than saying it has built a frontier coding assistant.

The New Bottleneck Is Human Judgment

Perhaps the most important implication of agent-directed model development is the increasing value of technical judgment.

AI agents can generate code quickly.

But generating code and making the correct engineering decision are different problems.

In large-scale training, an incorrect configuration can waste enormous computational resources.

A subtle data processing error can contaminate an evaluation.

A distributed training bug can remain hidden until thousands of GPU hours have been consumed.

An apparent performance improvement may disappear when tested against a more representative dataset.

The person directing the agents must understand these risks.

They need enough knowledge to challenge outputs, interpret metrics, and recognize when automated systems are producing plausible but incorrect results.

This creates an interesting paradox.

As AI agents reduce the need for routine implementation work, deep technical expertise may become more important for the people supervising them.

The most valuable researcher may not be the person who writes the largest amount of code.

It may be the person who consistently makes the best decisions about what code should be written and how its results should be evaluated.

What This Means for Software Engineers

The implications extend beyond AI research laboratories.

Software development teams are already experimenting with coding agents that can implement features, generate tests, review changes, and investigate bugs, as described in reports from Anthropic and OpenAI.

Mirendil's reported workflow represents a more ambitious version of the same idea.

For a broader look at how cheaper software creation is reshaping where value lies, read AI Is Making Software Cheaper to Build. So What Is Actually Becoming More Valuable?.

Instead of using agents to accelerate individual programming tasks, the researcher reportedly used them to support an entire complex engineering effort.

For software engineers, this suggests that several capabilities could become increasingly valuable.

1. System Architecture

Understanding how components interact becomes more important when agents can implement individual components quickly.

2. Verification and Testing

Generated code still needs validation. Engineers who can design reliable tests and evaluations will remain essential.

3. Technical Decision-Making

Choosing the right tradeoffs between performance, reliability, cost, and complexity becomes a central skill.

4. Agent Orchestration

Developers may increasingly need to coordinate multiple automated workers, define responsibilities, and manage shared context.

For a practical example of building and orchestrating AI agents, see Lets look at Google Agent Development Kit (ADK).

5. Domain Expertise

Knowing what should be built, why it matters, and whether the result solves the intended problem becomes a differentiator.

The lesson is not that software engineers will disappear.

It is that the relationship between engineering effort and engineering output may change substantially.

Could This Democratize Foundation Model Development?

If Mirendil's approach proves reproducible, it could lower one of the barriers to entering the foundation model market.

Smaller research organizations might be able to train specialized models without building large engineering departments.

Universities, independent research groups, and well-funded startups could potentially experiment with architectures and training methods that previously appeared organizationally impractical.

But democratization has limits.

Access to high-performance compute remains uneven.

High-quality datasets can be difficult to obtain or prepare.

Evaluating large models requires technical expertise and additional resources.

And competing with the largest AI laboratories on general-purpose capabilities may still require enormous budgets.

Agent-directed development could lower the organizational barrier without eliminating the financial barrier.

It might produce a more diverse ecosystem of specialized models rather than making frontier-scale general-purpose training accessible to everyone.

That would still be a meaningful change.

The $200 Million Question

Mirendil has reportedly raised a $200 million seed round led by Andreessen Horowitz and Kleiner Perkins.

That financing introduces an interesting contrast.

The company claims to have completed its initial pretraining effort for under $10 million, yet it has raised substantially more capital.

The difference suggests a larger ambition, although the funding amount alone does not establish how the company intends to spend the money.

Potential investments could include additional compute, new training runs, improved datasets, post-training, evaluations, infrastructure, and recruitment.

The broader question is whether the next generation of AI laboratories will use capital differently.

Instead of employing enormous teams to build every part of the training stack, they may invest more heavily in compute and agent infrastructure while maintaining smaller research organizations.

If this approach succeeds, the competitive advantage of an AI laboratory could increasingly depend on how effectively a small number of researchers direct enormous amounts of automated work.

What Still Needs to Be Proven?

Mirendil's announcement is promising, but several important questions remain unanswered.

Can the reported results be reproduced independently?

How does Mirendil-Base perform on standardized coding and reasoning benchmarks?

What exactly is included in the reported training cost?

How much human intervention was required when agents encountered difficult engineering problems?

Would the same approach work for multimodal models or substantially larger training runs?

Can the model be turned into a reliable instruction-following assistant through post-training?

These questions matter because a compelling engineering demonstration is not automatically proof of a broadly applicable industry transformation.

The distinction between a successful individual experiment and a repeatable methodology is fundamental to scientific progress.

Final Thoughts

The most important idea behind Mirendil's announcement is not that one person can replace an entire AI research laboratory.

It is that AI agents may allow a single researcher to coordinate engineering work that previously required many specialists.

That possibility changes how we think about the economics of building foundation models.

Compute remains expensive. Data remains essential. Research judgment remains difficult to automate.

But if implementation becomes dramatically cheaper, the balance between human labor and computational resources could shift.

The next generation of AI laboratories may compete not only on model architectures and training budgets, but also on how effectively their researchers direct automated engineering systems.

Mirendil has not yet demonstrated broad model superiority through independent benchmarks.

What it has presented is a provocative claim about how a large model can be built.

And that may be the more consequential story.

The future of AI research might not be defined solely by how many researchers a laboratory employs.

It may also be defined by how much meaningful research each person can direct.

TL;DR

• Mirendil claims one researcher directed AI agents to train Mirendil-Base from scratch in approximately 3.5 months.
• The model reportedly contains between 100 billion and 500 billion parameters.
• AI agents reportedly wrote more than 99% of the project's code.
• The company reports a pretraining cost below $10 million and says it used public data without copying existing models' outputs.
• Mirendil reports outperforming Nemotron-550B-Base and DeepSeek-V4-Pro-Base on its held-out code prediction metric, while trailing Kimi-K3.
• These results have not established superiority on general coding tasks or independent comprehensive benchmarks.
• Agent-directed development could reduce engineering headcount requirements while increasing the importance of compute, data, and research judgment.
• The biggest potential shift is organizational: smaller teams may be able to coordinate increasingly complex AI research projects.

References

• Mirendil — Pretraining at Scale Without a Pretraining Team
• OpenAI — Evaluating large language models trained on code
• SWE-bench — Benchmarking software engineering agents