All projects below are part of the Swallow Project, an initiative to build high-quality open Japanese-English bilingual LLMs for the research community.
Active Projects
Qwen3-Swallow
Status: In Development
Japanese-English continual pre-training, SFT, and RLVR on top of the Qwen3 base model. The goal is to achieve Japanese language capabilities that surpass Qwen3, while retaining English performance on par with Qwen3. By leveraging the Swallow Project’s data and training expertise, we aim to demonstrate that our post-training pipeline can match the quality of Qwen3’s own post-training.
GPT-OSS-Swallow
Status: In Development
Starting from GPT-OSS models that have already undergone SFT and RLVR, this project investigates how to enhance Japanese language knowledge and capabilities without degrading existing strengths in general dialogue, reasoning, and English proficiency. The goal is to build a strong Japanese-capable model while preserving the original model’s broad competencies.