Skip to content
Trending storyDeveloping

Mirendil announces Mirendil-Base, a coding model pretrained by one researcher with agents

1 article1 sourcesince Oct 21Last article 1h ago ·

Overview

AISummary of 1 article

Mirendil, a lab working on recursive self-improvement, announces Mirendil-Base, a coding model that one researcher pretrained from scratch in 3.5 months for less than $10M in compute.

Mirendil says agents wrote over 99% of the code.

On held-out code from Mirendil repositories, Mirendil-Base scores 0.280 bits per byte, ahead of GLM-5.3 and DeepSeek-V4-Pro-Base. Kimi-K3 still leads on outside repos at 0.193. The model is not released, and no weights, API, or task benchmarks such as SWE-bench have been shared, and post-training is still underway.

Written by AI from the articles below · updated Oct 21, 1:44 PM ET

Check the sources:

Latest developmentOct 21, 1:29 PM ET
Mirendil announces Mirendil-Base, a coding model pretrained in 3.5 months for under $10M

Article timeline

The articles in this story. Times are ET.

Oct 21
  1. Sophia YangXPick
    Mirendil announces Mirendil-Base, a coding model pretrained in 3.5 months for under $10M

    AIMirendil, a lab working on recursive self-improvement, announced its first model, Mirendil-Base. One researcher pretrained it from scratch in 3.5 months with less than $10M in compute, and agents wrote over 99% of the code. On held-out code from Mirendil repositories, it scores 0.280 bits per byte, beating GLM-5.3 and DeepSeek-V4-Pro-Base, while Kimi-K3 still leads at 0.193 on outside repos. The model is not yet released, and no weights, API, or task benchmarks such as SWE-bench have been shared.

Heat Trend

Not enough continuous observations to show a trend yet.