Eschatia Labs is an independent non-profit open-source AI research lab. We publish small AI models, data and papers that
people can run and test on their own computers. Our long-term goal is superintelligence.
| Work |
What it does |
Result |
| JevAlt decision models |
Deem-4B, Karar-4B and Wähler-4B choose between options in English, Turkish and German and give a probability for each |
Karar-4B, one of the best open Turkish decision models at 4B: 96.8% against Kev-4B's 87.1% on held-out Turkish decisions |
| Tholos-2B |
A small model that picks tools for AI assistants on your own computer |
137 of 160 Tholos-Bench scenarios against MiniCPM5-2B's 112 |
| xdfdet |
Eight deepfake video detectors with heatmaps that show which face regions drive each decision |
0.898 AUC on FaceForensics++ for the released detector |
| Eldalambë |
One of the best research efforts on language preservation with small local models |
Demo by invitation |
Open benchmarks (collection)
| Benchmark |
What it tests |
Size |
Links |
| JevAlt benchmark |
Choosing between options and giving a probability, in English, Turkish and German |
3,116 test rows |
dataset · page |
| Tholos-Bench |
Whether a small agent leaves a shared workspace of tables, notes and tasks in the right state |
160 scenarios in 14 categories |
dataset · page |
All open models and data: collection ·
GGUF builds for llama.cpp and Ollama: collection ·
eschatialabs.com · Mert Kaya · GitHub
