Shisa V2 Llama 3.3 70B is a bilingual chat model built on Meta's Llama-3.3-70B-Instruct foundation, designed to prioritize Japanese language performance while retaining strong English capabilities. Unlike earlier Shisa releases that relied on tokenizer modifications or extended pretraining, this version takes a leaner approach, focusing its optimization entirely on post-training. The model is intended for high-performance bilingual chat, instruction following, and translation tasks across Japanese and English, with native-level fluency as the north star for its Japanese output.

The training recipe centers on a refined mix of supervised fine-tuning and Direct Preference Optimization datasets, encompassing regenerated ShareGPT-style data, translation tasks, roleplaying conversations, and instruction-following prompts. This combination enables the model to handle diverse conversational styles while excelling at structured tasks. On Japanese benchmarks including JA MT Bench, ELYZA 100, and Rakuda, the model achieves leading task performance, demonstrating its strength in nuanced Japanese understanding. It inherits safety characteristics from its base model without additional alignment, and integrates smoothly with inference frameworks like vLLM and SGLang, making it practical for production bilingual applications that demand high-context understanding.