Hugging Face has released Luth-2-0.8B and Luth-2-2B, two large-scale language models that demonstrate advanced capabilities in French. These models are the result of significant improvements over their predecessor, Luth, and are trained on a diverse range of domains, including mathematics, code, and instruction following. The new models are post-trained from the Qwen3.5 family and incorporate reinforcement learning and multi-domain on-policy distillation (MOPD) to enhance their French language performance. They are designed to be lightweight and suitable for on-device use. The models are available for download along with the data, code, and evaluation setup. Source: huggingface
The release includes a broader set of benchmarks and evaluation changes, with a focus on improving the accuracy of French language tasks. For mathematics, the models use MGSM-Rev2, a corrected version of MGSM, with enhanced parsing and normalization for French decimal separators. In knowledge and science, the models use French subsets of standard multilingual MCQ benchmarks, with scores calculated in generation mode rather than using a log-probability metric. Code and instruction following benchmarks are also translated and verified for accuracy. The evaluation process ensures that all results can be reproduced using the provided setup. Source: huggingface
The models outperform every model in their size class across the selected benchmarks, while remaining competitive with models two to three times their size. Luth-2-0.8B surpasses Luth-0.6B-Instruct and ranks first among all tested models in its size class on each of the 12 benchmarks. Luth-2-2B follows a similar pattern, outperforming Luth-1.7B-Instruct on 11 of the 12 benchmarks and ranking first overall on 10. The evaluation results are based on non-reasoning mode and can be reproduced using the provided setup. Source: huggingface
