Sakana AI Launches Sakana Namazu, a Japanese-Specialized LLM API

Sakana AI has taken the model quietly running behind its Sakana Chat product and turned it into a standalone API called Sakana Namazu, live starting today. The pitch is narrow and deliberate: this is not another general-purpose chatbot trying to do everything in every language. It's built specifically for Japanese text and Japanese business habits, with web search and code execution baked in as native tools, and it drops into any OpenAI-compatible codebase with little more than a base_url swap.

Under the hood sits Kimi K2.6, the open model Moonshot AI released, which Sakana then fine-tuned on its own in-house data. The tuning wasn't just about vocabulary and grammar. Sakana says it specifically worked to cut down on the model dodging certain topics and to reduce bias in its answers, on top of sharpening how it handles Japanese-specific business context. The result, according to Sakana, keeps Kimi K2.6's reasoning strength intact while making the model noticeably sharper on anything distinctly Japanese.

The numbers back that framing up. On general reasoning benchmarks like AIME26, MMLU-Pro, and LiveCodeBench v6, Namazu holds onto the base model's performance rather than trading it away for localization. But the more telling jump shows up on FairPoliticsQA, a Sakana-built benchmark for political neutrality, where the score climbed from 34.10% to 56.30%. That's not a rounding-error improvement. It suggests the fine-tuning process meaningfully changed how the model handles culturally loaded questions, not just how fluently it writes.

Sakana is positioning Namazu as the answer to a gap it says Japanese companies keep complaining about: frontier APIs cost too much for everyday use, and dropping a raw open model into production leaves lingering doubts about output quality and data handling. Sakana's answer is a mid-tier option, priced to be usable at scale, wrapped in a familiar OpenAI-style API. The company is showcasing three use cases to make the point concrete — a fully autonomous weekly market research report that plans its own searches and writes itself, a customer-support and sales-data pipeline meant to consolidate scattered AI spending across a company, and, more whimsically, a fish tank light show where the model decides on visual themes, searches for reference images, and choreographs roughly a thousand simulated fish using function calling and image recognition of tank screenshots.

Enterprise deployment is next on Sakana's roadmap, and the company is explicit about crediting Moonshot AI's open release of Kimi K2.6 as the foundation none of this would exist without.