About
A personal benchmark project for text models and model generation.
Leberkäsbench is a personal project for comparing text models on small model-generation tasks. Most runs ask a model to generate a complete artifact, such as a single-file HTML canvas scene, an interactive HTML page, or a short QA answer set.
It is not trying to be a scientific benchmark. There are no scientific claims, no careful lab conditions, and no claim to completeness. It is a practical, curious, privately paid dashboard for seeing how different models behave on the same prompts.
How It Started
The whole thing started as a benchmark because I was curious what different models would do with the same tiny visual generation task. The other benchmark files came later, mostly because it was fun to compare more places, foods, maps, dialect tasks, and QA traps.
What It Measures
The dashboard looks at things that are useful to me when comparing models:
- whether the generated file is usable
- how fast the model starts and finishes
- how much the run costs
- how many tokens it uses
- whether QA-style answers match the expected answer key
For visual or interactive generations, the output is still judged with a lot of human taste involved. A model can be fast and cheap while making something boring, broken, or weird.
Maintenance
The runs are privately paid. I try to update the benchmark with new models as fast as possible, but availability, pricing, provider changes, and plain free time decide the actual pace.
If a model is missing, stale, or newly released, it is usually on the mental todo list already.
If you want to help me out with or sponsor the , feel free to reach out
Contact: atilla [at] atiuss [dot] com