> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Mirendil says one researcher and AI agents pretrained a large coding model from scratch
- URL: https://globalfeed.ai/en/mirendil-says-one-researcher-and-ai-agents-pretrained-a-large-coding-model-from-scratch/
- Published: 2026-10-23T02:30:53.000Z
- Updated: 2026-10-23T02:30:53.000Z
- Description: AI startup Mirendil says it pretrained its coding-focused Mirendil-Base model from scratch with AI agents directed by a single researcher, in 3.5 months and for under $10 million in compute. According to the company, agents wrote more than 99 percent of the code.
- Author: GlobalFeed Editor
- Tags: Mirendil, pretraining, AI agents, coding model, AI, dil-en, elle, x-mirendil

AI research startup Mirendil announced that it pretrained a large language model from scratch without a dedicated pretraining team, using AI agents directed by a single researcher. According to [the company's October 21 post](https://mirendil.com/news/pretraining-at-scale-without-a-pretraining-team/?ref=globalfeed.ai), the coding-focused Mirendil-Base model was trained with one researcher, in 3.5 months and for under $10 million in compute all-in. That figure includes the agents' token spend, data processing and small-scale experiments.

## Who Mirendil is

Mirendil announced its founding in June 2026\. According to [the launch announcement](https://mirendil.com/news/announcing-mirendil/?ref=globalfeed.ai), its goal is to accelerate science and technology by democratizing frontier AI research and development. The team consists of 20 researchers and engineers from institutions including Anthropic, xAI, Google DeepMind and OpenAI. Andreessen Horowitz and Kleiner Perkins led a $200 million seed round, and NVIDIA is also among the investors. Co-founder Behnam Neyshabur shared the new result on X.

## What was done

According to the announcement, AI agents wrote over 99 percent of the code, ran experiments and diagnosed failures, while the researcher decided what to try and what to trust. Starting from an empty repository, the agents built the training infrastructure, the model architecture, the training algorithm and the data pipelines, using generic libraries such as PyTorch and JAX and open-source projects as references. The company says efficiency came from architecture changes, data-mixture tuning and kernel optimization. The post does not say which models the agents ran on.

Training data consists of GitHub and other publicly available code, web pages and public datasets. The company says it used no purchased or licensed data, no distillation and no synthetic data.

## The model family

- **Mirendil-Base:** the smallest in the family, between 100 and 500 billion parameters; training is complete.
- **Mirendil-Base-Large:** between 500 billion and 2 trillion parameters, about 80 percent through training.
- **Mirendil-Base-XL:** between 2 and 10 trillion parameters, which the company calls a frontier-class model; about 19 percent through training.

## How it was measured

The company measured the model by cross-entropy in bits per byte on code (lower is better). The first comparison used held-out code from Mirendil's own repositories. The stronger second test used seven external code repositories published after the models were trained. On that test, Moonshot AI's Kimi-K3 ranks first on average and Mirendil-Base second, ahead of NVIDIA's Nemotron-550B-Base, Z.ai's GLM-5.3 and DeepSeek's V4-Pro-Base.

The comparisons have limits. Kimi-K3, at 2.8 trillion total parameters, is far larger than Mirendil-Base. As the company notes, Kimi-K3 and GLM-5.3 are post-trained models that score slightly worse than base models on this metric, so the comparison is not like-for-like. The metric shows how well a model predicts code; it does not directly measure practical success on coding tasks.

## Self-improving AI

The company's real goal is recursive self-improvement, AI building the next AI. Mirendil says its next model will be built partly by Mirendil models. It also states plainly that fully autonomous self-improvement has not been reached yet, and that Mirendil-Base did not build itself: an expert built it with AI tools.

Post-training of Mirendil-Base has begun; the company has not announced a release of model weights or an API for now. The cost, time and performance figures rest on the company's own statements and have not been independently verified, and the claim that agents wrote the code cannot be checked from outside. Large-model training becoming feasible for small teams would be an important signal for the industry, but the results need confirmation through independent testing. Follow assessments from independent researchers for other perspectives.