Qwenfant-0.4B

Behold, Qwenfant-0.4B, a Qwen3-based model trained for BabyLM 2026's Multilingual track. It has been trained on our half-synthetic dataset Qwenfood. Notably, it is also a character-level model, which we found to perform better than subword-level models.

Model Sources

• Repository: TBA upon release
• Paper: TBA upon release
• Dataset: leukas/Qwenfood