Sakana AI Launches Japanese LLM API

Sakana AI has launched Sakana Namazu, an API service that makes its specialized Japanese large language model available to developers and businesses. The model, previously powering Sakana Chat, has been enhanced for improved Japanese language understanding and agentic capabilities. This move addresses a growing demand for robust, Japanese-centric AI tools in the market.

The Sakana Namazu API is based on Moonshot AI's open-source Kimi K2.6 model. Sakana AI has refined this base model using proprietary internal data, focusing on adapting it to the nuances of the Japanese language and specific business practices within Japan. Additionally, the company has implemented tuning to reduce unwanted responses and minimize output bias. This iterative development builds upon the foundation laid by the initial Sakana Chat release.

Addressing a Market Gap

The release comes at a time when options for Japanese-language LLM APIs suitable for professional use remain limited. High-tier models often come with significant costs, while relying on general open-source models can raise concerns about quality, data handling, and cultural context. Sakana AI positions Sakana Namazu as a solution that bridges this gap, offering a tailored Japanese LLM at a competitive price point with an accessible API.

StartupHub.ai data indicates that Sakana AI, which has raised $19M in 2024, scores a 70/100, placing it in a competitive segment. Competitors like Adept AI (71/100) and Mistral AI (69/100) also focus on advanced AI capabilities, highlighting the dynamic nature of the LLM development space.

Enhanced Capabilities and Benchmarks

Sakana Namazu demonstrates strengths in both advanced reasoning and Japan-specific performance. The model maintains the high capabilities of its Kimi K2.6 base, showing strong results in benchmarks for mathematical reasoning (AIME26), broad knowledge (MMLU-Pro), and coding (LiveCodeBench v6). This indicates a capacity for complex problem-solving.

Crucially, the model shows marked improvement in Japan-centric tasks. It surpassed its base model in JFBench, a Japanese instruction-following benchmark, and in internal benchmarks for Japanese-to-English translation that account for honorifics and Japan-specific nouns. A significant leap was observed in FairPoliticsQA, a neutrality test, where performance increased from 34.10% to 56.30%. This suggests a deeper comprehension of Japanese cultural context and values.

Practical Applications

The API is designed for a range of practical applications. One described use case is the automated generation of market research reports, handling tasks from planning and web searching to data synthesis and final report writing. This relies on the model's inherent agentic abilities and integrated web search tools.

Another application involves automating customer support and analyzing sales data, leveraging the model's specialized Japanese language skills for end-to-end processing. The API's cost-effectiveness is highlighted here, enabling broader adoption across an organization.

A more creative demonstration involves an automated fish light show. Given a theme like 'the world of the ocean,' the model autonomously plans sequences, searches for visual references, and generates instructions for coordinating a school of fish. This showcases the model's agentic capabilities, combined with function calling and image recognition, extending beyond typical business automation into creative domains.

Industry Context and Future Outlook

Sakana Namazu's development is rooted in the open AI community, particularly acknowledging Moonshot AI's contributions. The company also signals future enterprise-focused deployments and is actively recruiting talent for its expansion in Japan.