Luth-2: New State-of-the-Art French Small Language Models Released

The Luth-2 project has released two new small language models, Luth-2-0.8B and Luth-2-2B, which establish a new state-of-the-art for French language tasks. These models leverage the Qwen3.5 architecture as their backbone and utilize advanced post-training techniques like multi-domain on-policy distillation. These models demonstrate that small, local-friendly models can outperform much larger models (up to three times their size) on non-English tasks when properly optimized. This highlights a significant untapped potential in tailoring small language models for high-resource languages other than English. Luth-2 was trained on a new 3-billion-token SFT mixture covering domains like coding, math, and tool calling, combined with reinforcement learning. In benchmarks, Luth-2-2B scored 69.67 on Multi-IF and 81.52 on Math-500, outperforming Gemma-4-E2B-it.