Shisa AI Launches Shisa V2: Next-Level Bilingual LLM for Japanese Tasks Goes Free

Shisa V2 Series Official Release: A Breakthrough in Bilingual LLM Technology

Shisa.AI, headquartered in Minato, Tokyo, has announced the official launch of the Shisa V2 series, a cutting-edge bilingual chat model designed for Japanese and English tasks. This innovative model boasts exceptional performance in Japanese language processing as verified by industry-standard benchmarks, while also maintaining top-tier capabilities in English. Notably, the series adds new 12B and 14B parameter models optimized for cost-performance in real-world applications. All model sizes, including 7B, 8B, 12B, 14B, 32B, and 70B, are now freely available under open licenses such as Apache 2.0 and MIT on HuggingFace.

Key Enhancements

1. Significant Improvement in Japanese Processing Capability
The Shisa V2 model builds upon the success of Shisa V1, which achieved over one million downloads. The new version has outperformed multiple model classes in industry-standard benchmarks. By using synthetic data for training optimization instead of relying solely on costly pre-training methods, Shisa V2 has achieved impressive advancements in:
- Naturalness of output and contextual understanding
- Performance in practical tasks, such as translation and role-playing

2. Expanded Model Family
Shisa V2 encompasses a broad range of versions from the compact 7B model to the large 70B model. It significantly enhances Japanese processing capabilities over baseline models like Qwen2.5 and Llama3.1, demonstrating high practicality across various scenarios from everyday tasks to specialized applications.

License	Model Name	Parameters	Japanese Benchmark	English Benchmark
----	-------	--	----	------
Apache 2.0	shisa-v2-qwen2.5-7b	7B	71.06	54.86
Llama 3.1	shisa-v2-llama3.1-8b1	8B	70.83	54.75
Apache 2.0	shisa-v2-mistral-nemo-12b	12B	72.83	53.33
MIT	shisa-v2-unphi4-14b	14B	75.89	60.10
Apache 2.0	shisa-v2-qwen2.5-32b	32B	76.97	67.41
Llama 3.3	shisa-v2-llama3.3-70b1	70B	79.72	67.71

Newly Developed Japanese Benchmarks

Shisa V2 not only excels in conventional performance metrics but also targets practical tasks that were previously challenging to measure. The company has developed proprietary Japanese benchmarks that evaluate:

- shisa-jp-ifeval: Advanced Japanese instruction following ability
- shisa-jp-rp-bench: Complex role-play and multi-turn dialogue capabilities
- shisa-jp-tl-bench: High-quality Japanese-English translation performance

These benchmarks will soon be made open-source to contribute to the development of a diverse AI research community in Japan.

About Shisa.AI

Shisa.AI is a next-generation AI startup led by a team rooted in Silicon Valley's cutting-edge technology. The company merges advanced technology from Silicon Valley with deep insights into the Japanese market, focusing on two main pillars: "Japanese-specific AI" and "data-driven development." Shisa.AI is committed to advancing Japanese language processing technologies and promoting Japanese-origin AI innovations to the global market.

Strategic Partnership with AKA Virtual

In recognition of Shisa.AI's superior Japanese LLM technology, AKA Virtual established a strategic partnership in October 2024. This collaboration involves integrating Shisa.AI's large language model (LLM) technology into AKA's virtual idol business. Notably, in AKA Virtual's AI character service “DE-AI,” Shisa's advanced LLM technology serves as the core system, delivering a high-quality user experience.

For more information, you can visit the official websites: