What is OpenMythos? A Simple Guide to the New AI Architecture

What is OpenMythos? A Simple Guide to the AI Architecture

If you follow artificial intelligence, you know that tech companies keep their best secrets locked up. We know models like Claude are incredibly smart, but we do not officially know exactly how they work under the hood.

Recently, a developer named Kye Gomez created an open-source project called OpenMythos. It is a theoretical reconstruction of how the new Claude Mythos model probably functions. Gomez built it from scratch using clues from published research papers.

The main takeaway? The future of AI isn’t just about building bigger brains. It is about building brains that recycle their own thoughts.

Here is a breakdown of how OpenMythos works, explained without the complicated math.

The Magic of the Loop

Most normal AI models operate like an assembly line. If the AI has 100 layers, a piece of text travels through layer 1, then layer 2, all the way to layer 100. Every single layer is a completely different set of code and parameters. This takes up a massive amount of memory.

OpenMythos uses a different structure called a Recurrent-Depth Transformer (sometimes called a Looped Transformer). It breaks the assembly line into three parts:

1. The Prelude: The starting point. The text goes through a few normal layers just once.
2. The Recurrent Block: The middle. This is where the loop happens.
3. The Coda: The finish line. The text goes through a final few layers before giving you an answer.

Instead of passing the text through 100 unique layers, OpenMythos passes the text through a small “Recurrent Block” and just loops it over and over again. It uses the exact same brainpower multiple times. More loops equal deeper thinking, without taking up more computer storage.

Thinking in Secret

Usually, if you ask an AI a really hard math problem, it has to write out its thought process to get the right answer. It literally types out on your screen, “First I will add the numbers, then I will multiply…” OpenMythos does not do that. It reasons silently.

Every time the model loops the information in its Recurrent Block, it acts like a step in a thought process. But it happens in a “continuous latent space”—which is a fancy way of saying the AI does the thinking completely inside its own head. By the time it actually starts typing out the answer, it has already explored different ideas, looped through the problem, and found the correct path.

How Does It Know Everything? (Mixture of Experts)

If the AI uses a small, recycled loop, how can it write poetry, write Python code, and pass a law exam all at the same time?

It uses a system called Mixture of Experts. Inside that looping middle block, the AI is split into tiny specialists. When a prompt comes in, a “router” decides which expert needs to look at it. If you ask a coding question, the math and poetry experts stay asleep, and only the coding expert turns on.

This means the overall OpenMythos model can store a massive amount of information, but it only activates about 5% of its brain at any given moment. It is fast, efficient, and cheap to run.

Why Does OpenMythos Matter?

Building huge AI models requires billions of dollars and giant warehouses full of computer chips.

The OpenMythos project proves that we might not need to keep making models bigger to make them smarter. A looped model with fewer parameters can achieve the exact same quality as a massive model, simply by looping its thoughts a few extra times.

It is a fascinating glimpse into where AI is headed. Instead of just throwing more raw power at a problem, developers are figuring out how to make AI stop, think, and loop before it speaks.