Luth-2 French SLM Released

French-specialized small language models Luth-2-0.8B and Luth-2-2B have been released. The developers stated that the two models achieved top performance on major French benchmarks targeting models of similar size.

• Multi-IF: Luth-2-2B 69.67 points, Gemma-4-E2B-it 65.17 points
• MGSM-Rev2: Luth-2-0.8B 72.92 points, granite-4.0-h-micro 55.60 points
• Math-500: Luth-2-2B 81.52 points, Gemma-4-E2B-it 81.24 points

This version used 3 billion tokens of SFT data covering math, knowledge, code, tool calling, instruction following, multi-turn dialogue, and science. It applied expert-specific reinforcement learning and multi-domain on-policy distillation (MOPD), and adopted Qwen3.5 as the backbone due to its high post-training adaptability.

The two models were released on Hugging Face targeting local and on-device execution, along with SFT and RL data, training code, and a French LLM leaderboard.