About
Shisa AI (the augmxnt project) builds bilingual Japanese–English chat models by post-training strong open base models on synthetic Japanese data, releasing the models, datasets, and training recipes openly. The Shisa V2 and V2.1 families span roughly 1.2B to 405B parameters across Llama, Qwen, and Mistral bases, and score at or near the top of Japanese benchmarks like JA MT-Bench for their size classes. It is the go-to open option when an application needs natural, culturally aware Japanese rather than translated English behavior.
Download the weights from the shisa-ai organization on Hugging Face and run them locally with Transformers, vLLM, or llama.cpp-compatible runtimes. Choose the size that fits your hardware — the 7B–8B models run comfortably on a single consumer GPU.