Blog

A year ago we said no open-source large language model could beat Claude at parsing German grammar, so we trained our own specialist instead. The satzklar model matches treebank state-of-the-art (97.7% attachment accuracy), runs in milliseconds on a plain CPU, and now powers grammar analysis on satzklar.net. Open-sourced on Hugging Face.

We tested GPT-OSS 120B, Llama 4, DeepSeek R1, and other open source models against Claude 4 for German parsing. Despite 100% parse success, GPT-OSS drops 50% of words. Some day an open source model will beat Claude. Not today.

German pronunciation is challenging for learners. This technical deep-dive explores our journey from unreliable browser TTS APIs to a robust Narakeet implementation that powers natural German audio throughout our application - from word pronunciation to story reading.

German learners face a unique challenge: the language's flexible word order and complex grammatical relationships create a web of dependencies that traditional learning tools struggle to illuminate. This post explores how we built satzklar.net entirely using Claude for both development and runtime language processing.