Shisa.AI Unveils Shisa V2.1: A Bilingual Model That Is Smaller, Smarter, and More Accessible

Shisa.AI Launches Shisa V2.1: A Leap Forward in Bilingual AI

Shisa.AI, based in Minato, Tokyo, has announced the release of its highly anticipated Shisa V2.1 series, scheduled for December 9, 2025. This new series introduces a set of bilingual chat models designed to significantly enhance performance in both Japanese and English, making them suitable for a variety of applications across different devices. From compact 1.2 billion parameter models suitable for smartphones to powerful models with 14 billion and 70 billion parameters that rival GPT-4 in Japanese performance, Shisa V2.1 comes in five different sizes. Additionally, for the first time, Shisa.AI will offer paid API access alongside free downloads from HuggingFace, expanding accessibility and usability for users worldwide.

Smaller and Smarter

The Shisa V2.1 series is distinguished by its compact design, which allows it to deliver superior Japanese performance despite its reduced model size. The 14B variant, for instance, surpasses the Japanese language capabilities of the previously released Shisa V2 70B model, achieving exceptional results at a fraction of the size. Furthermore, the 70B model reaches near the performance levels of the flagship Shisa V2 405B, also significantly smaller in terms of resource consumption compared to its predecessor. This efficiency translates into reduced memory requirements and faster inference speeds, making it more cost-effective and accessible for a broader range of applications and environments.

Performance Metrics

The release documents provide a comprehensive comparison of language capabilities across various models and their parameters:

Model Name	Parameters	Japanese AVG	English AVG	JA-MT
-----	--	---	----	-
Llama 3.1 (Shisa V2 405B)	405B	74.7	67.5	9.43
Llama 3.3 (Shisa V2.1 70B)	70B	73.1	66.0	9.26
Shisa V2.1 14B	14B	72.6	57.7	9.28
Shisa V2.1 8B	8B	67.8	57.8	8.93
Shisa V2.1 3B	3B	57.9	43.2	7.55
Shisa V2.1 1.2B	1.2B	43.4	27.6	6.69

These enhancements are attributed to a revitalized dataset and the integration of advanced techniques, including Reinforcement Learning (RL) and model merging technologies. Rather than focusing entirely on benchmark training, the results reflect real-world improvements in Japanese language performance.

Accurate Japanese Output

As the number of Japanese-capable LLMs expands rapidly, Shisa.AI recognizes the injection of foreign languages into Japanese outputs as a critical problem. To address this, the company has pioneered a quantification technique, termed Cross-Lingual Token Leakage, which evaluates the leakage of foreign language tokens during output generation. The results have been promising, achieving a leakage reduction rate that significantly surpasses other models:

Base Model	Base Leakage Rate	V2.1 Leakage Rate	Improvement Factor
-----	---	-----	----
Llama 3.2 3B	11.48%	0.24%	47.8×
LFM2 1.2B	4.32%	0.32%	13.5×
Qwen 3 8B	2.18%	0.44%	5.0×
Llama 3.3 70B	1.90%	0.36%	5.3×
Phi 4 14B	0.12%	0.06%	2.0×

The inclusion of mixed languages in actual applications, such as translation and customer support, can derail functionality. By achieving maximum improvements of up to 47.8 times compared to baseline models, Shisa V2.1 is positioned as a market leader in this area.

API Launch and Commercial Services

In conjunction with the Shisa V2.1 release, Shisa.AI is also rolling out text, translation, and speech APIs. Their translation service, “chotto.chat,” is already live, offering high-quality Japanese-English translations for both individuals and businesses. This approach not only addresses the high barriers to testing performance-intensive LLMs, such as cost and infrastructure setup, but also promises accessible pricing through OpenRouter, including free tier options. Custom learning, dedicated capacity, and on-premises deployment are also available to meet specific business needs.

The Shisa V2 models have received recognition from Japan’s Ministry of Economy, Trade, and Industry as part of the GENIAC initiative for domestic models. All computation resources are hosted within Japan, allowing for low-latency performance and compliance with data residency and regulatory requirements.

Learning on AMD Hardware

Training for the Shisa V2.1 model was conducted on AMD MI300X GPUs provided by the AMD Developer Cloud. This marks the first large-scale LLM developed in Japan to leverage AMD hardware, which has been further optimized through proprietary modifications that are also being open-sourced, establishing a higher standard for future model training efficiency.

About Shisa.AI

Founded by a team of tech innovators based in Silicon Valley, Shisa.AI is a next-generation AI startup focused on developing Japanese language AI capabilities and data-driven methodologies. The company is at the forefront of evolution in open-source LLMs, striving to share Japanese AI innovations on a global scale.

Related Links

For media inquiries, please contact: Shisa.AI Public Relations at [email protected]