Intel / article

Sakana AI ships Namazu, a Japanese-tuned LLM API on Kimi K2.6

Source: sakana.ai
Author: Sakana AI
Date: Published Aug 3, 2026

Context

Sakana Namazu is a Japanese- that Sakana AI sells through an API, and it was not trained from scratch. It starts from Kimi K2.6, an open model released by Moonshot AI, which Sakana further trained on its own data for Japanese and for Japanese business work. According to Sakana's API announcement, that training also aimed to reduce unnecessary refusals on certain topics and to limit bias in its answers. The product page pitches the result for business documents and email in natural Japanese, including keigo, the honorific register used in business correspondence.

Sakana's case for choosing Namazu over its base model rests on its own comparisons. The announcement says Namazu kept Kimi K2.6's performance on math, general-knowledge and coding tests (AIME26, MMLU-Pro and LiveCodeBench v6) while beating it on three Japanese-focused measures: JFBench, a Preferred Networks for following instructions in Japanese; a Japanese-to-English translation task; and FairPoliticsQA, which checks whether answers stay neutral rather than leaning toward one country's values. The only score Sakana gives in its text is for FairPoliticsQA, which rose from 34.10% to 56.30%. Sakana built the translation task and FairPoliticsQA itself, so its headline improvement comes from a test the company designed.

Switching is meant to take a configuration change. The API is compatible with OpenAI's, so existing code points its base URL at Sakana and sets the model to sakana-namazu, and web search and code execution are built into the same API, so the model can look things up and run calculations within a single request. Pricing is $0.95 per million input and $4.00 per million output tokens, plus $7 per 1,000 web searches, and Sakana also lists Namazu on OpenRouter. Two defaults matter for a company weighing it: inputs may be used to train Sakana's models unless the customer opts out in the console, and Sakana does not guarantee that processing stays inside Japan. The service is not yet available in the EU, the EEA, the UK or Switzerland.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source sakana.ai