Shisa V2 70B

Shisa V2 70B (shisa-ai/shisa-v2-llama3.3-70b) is a bilingual Japanese-English chat model fine-tuned by Shisa.AI on Meta's Llama-3.3-70B-Instruct base. Unlike earlier Shisa releases it skips tokenizer surgery and continued pretraining entirely, focusing purely on post-training: a refined mix of supervised fine-tuning and DPO over regenerated ShareGPT-style data, translation tasks, roleplay conversations and instruction-following prompts. The team reports leading Japanese task performance across custom and public benchmarks (JA MT Bench, ELYZA 100, Rakuda) while retaining strong English, with a 128K context and clean vLLM/SGLang compatibility.

It inherits the Llama 3.3 Community License from its base, and no additional safety alignment was applied beyond that. Shisa.AI later extended the family up to a 405B model, which it calls the highest-performing LLM ever trained in Japan. Date is era-level based on repository history; verify the exact release if precision matters.