# Aleph Alpha trains a model to reason in German

> Aleph Alpha says roughly 800,000 German training examples helped a small AI model produce German reasoning traces. The work also exposed a trade-off: German benchmark scores initially fell, partly because the model got stuck repeating its thoughts.

Oossa · 2026-09-29 · https://oossa.com/en/aleph-alpha-trains-a-model-to-reason-in-german

Aleph Alpha researchers Niklas Finken and Simon Thel describe a method for teaching an AI model to reason in German, not just answer in German. They generated about 800,000 training examples by prompting other models to begin their reasoning in German, then used the examples to fine-tune a model with 3 billion active parameters.

The company says the model’s English performance stayed unchanged, but German math scores first dropped: on German AIME 2026, the score fell from 70.2 to 48.3, then recovered to 67.3 in the strongest math-data run. The researchers say adding German data from the same subject area helped more than adding German data in general, and that the approach does not yet reliably stop the model from looping.

## The facts

- The researchers generated about 800,000 German supervised fine-tuning examples.
- German AIME 2026 score: 70.2 before the change, 48.3 at first, and 67.3 in the strongest math-data run.

## Why it matters

For German-speaking users, the work points toward AI reasoning they can read in their own language, but shows that improving language consistency can come with accuracy and reliability costs.

## Sources & references

1. [Through the Valley of Tears: Cold Starting German Reasoning in LLMs](https://www.aleph-alpha.com/en/blog/through-the-valley-of-tears-cold-starting-german-reasoning-in-llms/) – Aleph Alpha, 2026-09-29

Last updated: 2026-09-29