Source: https://eschatialabs.com/

![Eschatia Labs logo: a line drawing of Pegasus rearing above mountain peaks inside a dotted circle](https://eschatialabs.com/brand/logo-640.webp)

# An international open-source research lab.

We publish artificial intelligence models, data and papers so you can test their results and understand their decisions. Our long-term goal is superintelligence: systems that exceed human reasoning across fields.

[See the research](https://eschatialabs.com/research/)

## [Detecting AI-manipulated faces](https://eschatialabs.com/research/#deepfake-detection)

Detection scores need explanations. We compare training methods and map the facial regions behind predictions.

Detection score (AUC): 0.897 for EfficientNet-B4 with varied and black-masked training images, versus 0.868 for unmodified EfficientNet-B4, on the same deepfake video data (FaceForensics++).

[UBMK 2026 paper](https://eschatialabs.com/research/ubmk-2026/) [MSc thesis](https://eschatialabs.com/research/msc-thesis/)

- [xdfdet](https://xdfdet.mertkayacs.com/) Eight deepfake detectors, code, explanation maps and a game.
  - [Real or AI?](https://xdfdet.mertkayacs.com/game/)
  - [Code (GitHub)](https://github.com/mertkayacs/xdfdet)
  - [Models and data (Hugging Face)](https://huggingface.co/mertkayacs/xdfdet)

## [Decision models](https://eschatialabs.com/research/#decision-models)

We train small models to answer questions and estimate uncertainty in English, Turkish and German.

Karar-4B answers 96.8% correctly, versus 87.1% for [Kev-4B](https://huggingface.co/jaredpalmer/kev-4b), on the project’s own Turkish test data kept separate from training.

[Technical report](https://eschatialabs.com/research/jevalt-report/)

- [JevAlt](https://jevalt.mertkayacs.com/) Decision models for English, Turkish and German.
  - [Code (GitHub)](https://github.com/mertkayacs/jevalt)
  - [Models and data (Hugging Face)](https://huggingface.co/collections/mertkayacs/jevalt-6abde16559675245349acea0)
- [Emberwick](https://emberwick.mertkayacs.com/) A medieval village where the JevAlt models decide.

## [AI on your own computer](https://eschatialabs.com/research/#local-ai)

To keep data under your control, we build task assistants for your own computer.

Tholos-2B completes 137 of 160 workspace tests on graphics hardware (Kaggle T4), 25 more than its starting model MiniCPM5-2B. On a computer’s main processor (CPU): 134.

[Technical report](https://eschatialabs.com/research/tholos-2b-report/)

- [Tholos](https://tholos.mertkayacs.com/) Task assistants sharing tables, notes and a task board.
  - [Code (GitHub)](https://github.com/mertkayacs/tholos)
  - [Models and data (Hugging Face)](https://huggingface.co/mertkayacs/Tholos-2B)
- [reevesagents](https://reevesagents.mertkayacs.com/) Open source software development with AI coding assistants.
  - [Code (GitHub)](https://github.com/mertkayacs/reevesagents)
  - [Package (npm)](https://www.npmjs.com/package/reevesagents)

## [Language preservation](https://eschatialabs.com/research/#language-preservation)

One of the best research efforts on language preservation with small local models: we use dictionaries and grammar checks to help preserve languages with few written records.

Eldalambë reads and writes Tolkien’s invented Elvish languages (Quenya and Sindarin) and his Elvish script (Tengwar).

- [Eldalambë](https://eldalambe.mertkayacs.com/) Dictionary-checked translation. Private code and models; access by invitation.

[5,700+](https://huggingface.co/mertkayacs) downloads on Hugging Face and [3,400+](https://www.npmjs.com/package/reevesagents) on npm, as of October 7, 2026.

## Contact

Write to [Mert Kaya](https://mertkayacs.com/), who founded the lab.

[mertkayacs@gmail.com](mailto:mertkayacs@gmail.com)