lightonai/Qwen3-8B-DE

Overview

lightonai/Qwen3-8B-DE is an 8 billion parameter model from the Qwen3 family, developed by lightonai. It is uniquely fine-tuned to perform native reasoning in German, meaning it produces its entire chain-of-thought (CoT) and final answer in German. This model was released in conjunction with the paper "Rethinking the Multilingual Reasoning Gap with Layer Swap" and is part of a specialized German trio designed to investigate multilingual reasoning.

Key Capabilities & Features

• Native German Reasoning: Excels at generating detailed reasoning processes and answers entirely in German.
• Base Model: Fine-tuned from Qwen/Qwen3-8B-Base.
• Context Length: Supports a substantial context window of 32,768 tokens.
• Training: Underwent full Supervised Fine-Tuning (SFT) over approximately 10 billion tokens across 2 epochs, utilizing the German split of the lightonai/Dolci-Think-SFT-32B-Multilingual dataset.

Performance Highlights

Evaluated on German versions of various benchmarks, Qwen3-8B-DE demonstrates strong performance, achieving an average accuracy of 72.59% across MGSM-Rev2, Global-MMLU-Lite, GPQA-Diamond, AIME 24/25, and HumanEvalPlus. While other related models in the trio (like Qwen3-8B-DE-Pivot-EN) show higher average scores by pivoting to English CoT, Qwen3-8B-DE stands out for its dedicated native German reasoning capability.

Good for

• Applications requiring complex problem-solving and reasoning in German.
• Use cases where the entire thought process, not just the final answer, needs to be in German.
• Research into multilingual reasoning and the impact of native language CoT generation.