kurakurai/Luth-2-2B

Luth-2-2B: A State-of-the-Art French Language Model

Luth-2-2B, developed by kurakurai, is a 2.3 billion parameter model (1.88B text-only) specifically engineered to excel in French language tasks. Built upon the Qwen3.5-2B architecture, this model has achieved state-of-the-art performance for its size across various domains, making it suitable for efficient local and on-device deployment.

Key Capabilities & Training

• French Language Mastery: Luth-2-2B sets new benchmarks in French for math, code, instruction following, general knowledge, and tool calling, outperforming other models in its size class.
• Advanced Training Methodology: The model underwent a two-stage post-training process:
  • Supervised Fine-tuning (SFT): Trained on a 3 billion-token French mixture dataset, Luth-2-Post-Training-SFT, covering math (37.2%), knowledge (27.9%), code (22.2%), instruction following (6.5%), and tool calling (6.3%).
  • Multi-domain On-Policy Distillation (MOPD): Utilized three specialized models (math, code, instruction following) trained with GRPO on Luth-2-Post-Training-RL, then distilled back into the SFT student.
• Strong Benchmarks: Demonstrates superior performance on French benchmarks such as MGSM-rev2 (86.52), HumanEval+ (66.00), and IFEval (75.06), as detailed on the French LLM Leaderboard.

Use Cases

• French-centric Applications: Ideal for applications requiring high-quality French language understanding and generation.
• Resource-Efficient Deployment: Its compact size makes it suitable for local inference, on-device applications, and scenarios where computational resources are limited.
• Specialized Tasks: Excels in French math problem-solving, code generation, instruction following, and general knowledge queries.