What is Luth-0.6B-Instruct?

Luth-0.6B-Instruct is a 0.8 billion parameter instruction-tuned model developed by kurakurai, based on the Qwen3-0.6B architecture. It has been specifically fine-tuned using the Luth-SFT dataset to drastically enhance its French language proficiency across instruction following, mathematical reasoning, and general knowledge tasks. Notably, this specialization was achieved while preserving and even improving its English capabilities in certain areas.

Key Capabilities & Performance

• Bilingual Proficiency: Excels in both French and English, with significant improvements in French benchmarks.
• Instruction Following: Demonstrates strong instruction-following abilities in French.
• Mathematical & General Knowledge: Shows enhanced performance in French math and general knowledge tasks.
• Compact Size: At 0.8 billion parameters, it offers efficient performance for its size.

Benchmark results using LightEval show Luth-0.6B-Instruct outperforming Llama-3.2-1B and the base Qwen3-0.6B in most French benchmarks, including IFEval, GPQA-Diamond, MMLU, Math500, and Arc-Challenge. In English, it maintains competitive scores, even surpassing the base model in MMLU, Math500, and Arc-Challenge.

When to Use This Model

Luth-0.6B-Instruct is particularly well-suited for applications requiring a small, efficient language model with strong bilingual capabilities, especially for tasks involving French instruction following, content generation, or knowledge retrieval. Its balanced performance across both languages makes it a versatile choice for developers targeting a French-speaking audience while also needing solid English support.