Asagi

Open Japanese vision-language model family (2B/4B/8B/14B) from Tatsuya Harada's Machine Intelligence Lab, with RIKEN and NII contributors: the training data is synthesized from scratch using only open models — deliberately avoiding outputs of usage-restricted LLMs like GPT-4 — making the whole stack license-clean. Top score among Japanese VLMs free of such restricted data; Apache-2.0 checkpoints under the MIL-UT org. Base of the medical Med-Asagi line.