lightonai/Qwen3-8B-FR

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Mar 23, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

The lightonai/Qwen3-8B-FR model is an 8 billion parameter native reasoning model, fine-tuned from Qwen/Qwen3-8B-Base, specifically designed to perform its entire reasoning trace and deliver final answers in French. Trained on approximately 10 billion tokens over two epochs with a 32,768-token context length, it excels at complex problem-solving requiring French-native chain-of-thought. This model is optimized for French language tasks, particularly those benefiting from explicit reasoning steps.

Loading preview...

lightonai/Qwen3-8B-FR: French Native Reasoning Model

lightonai/Qwen3-8B-FR is an 8 billion parameter language model developed by Lighton AI, fine-tuned from Qwen/Qwen3-8B-Base. Its primary distinction is its capability to perform native reasoning entirely in French, generating a complete chain-of-thought (CoT) before providing the final answer, also in French. This model was released in conjunction with the research paper "Rethinking the Multilingual Reasoning Gap with Layer Swap."

Key Capabilities & Training

  • French-Native Reasoning: Produces both its reasoning trace and final answer exclusively in French.
  • Base Model: Built upon the robust Qwen/Qwen3-8B-Base architecture.
  • Training Data: Underwent full Supervised Fine-Tuning (SFT) on approximately 10 billion tokens over two epochs, utilizing the French split of the lightonai/Dolci-Think-SFT-32B-Multilingual dataset.
  • Context Length: Supports a substantial context window of 32,768 tokens.

Performance & Related Models

This model is part of a specialized trio investigating the multilingual reasoning gap. While Qwen3-8B-FR achieves an average accuracy of 72.36% across French benchmarks like MGSM-Rev2, Global-MMLU-Lite, GPQA-Diamond, AIME 24/25, and HumanEvalPlus, related models such as Qwen3-8B-FR-Pivot-EN (CoT in English) show higher average performance (75.04%), highlighting ongoing research into multilingual reasoning strategies.

Ideal Use Cases

  • Complex French Problem Solving: Suited for applications requiring detailed, step-by-step reasoning in French.
  • Research on Multilingual Reasoning: A valuable tool for studying and developing models with native language CoT capabilities.
  • French Language AI Applications: Any scenario where high-quality, explainable outputs in French are critical.

Finetunes
3 models