Open-weight models for German and European documents

Many model publishers release downloadable weights while retaining licence conditions that differ from conventional open-source software. Some permit broad commercial use; others add attribution, redistribution, acceptable-use or scale conditions.

For operational use, record the publisher, model name, exact revision, weight format, quantisation, licence URL and download source. ‘Llama’, ‘Mistral’ or ‘Qwen’ alone is not enough to reproduce a deployment.

A general multilingual claim does not establish reliable German legal phrasing, Swiss orthography, Austrian terminology or mixed-language correspondence. Evaluation prompts should preserve the vocabulary and document structure of the intended users.

Measure whether the model keeps defined terms stable, distinguishes quoted text from interpretation, preserves numbers and dates, follows output format and declines when the source does not answer the question.

Scanned documents need OCR. Large collections need chunking, embeddings and often reranking. Answers need source presentation. A weaker base model with a disciplined retrieval pipeline can outperform a larger model receiving incomplete or poorly extracted text.

This is why SelbsAI treats model choice as a workload package rather than a permanent brand promise. The model, retrieval components and appliance envelope need to be tested together.

A compact test set of 20 to 50 representative tasks is more useful than a vague claim that a model is ‘good at German’. Store the expected source, critical facts, acceptable output criteria and reviewer notes without turning confidential production files into uncontrolled test data.

Run the set after model, quantisation, prompt, OCR, embedding or runtime changes. Publish only results whose hardware, versions and scoring method can be inspected.