Sakana AI launches Namazu, a Japanese-tuned LLM API
Tokyo lab Sakana AI opened commercial access to Namazu, a Japanese-specialised model built by fine-tuning Moonshot's open-weight Kimi K2.6.
An open Chinese base, tuned for Japanese business
Sakana AI began offering Namazu, a Japanese-focused large language model API, on Monday. The Tokyo lab says the model starts from an open-weight base — Moonshot's Kimi K2.6 — and adds fine-tuning on Sakana's own data to fit Japanese language use and Japanese workplace context, including natural keigo and business convention. The tuning also targets output behaviour, suppressing bias and adding refusal handling on sensitive topics.
On benchmarks, Sakana reports that Namazu holds roughly the base model's level on mathematics (AIME26) and coding (LiveCodeBench v6) while improving substantially on Japanese-specific evaluations; on FairPoliticsQA, a political-neutrality test, the score rises from 34.10% to 56.30%.
Pricing and positioning
Namazu is sold pay-as-you-go with no monthly fee: $0.95 per million input tokens and $4.00 per million output tokens, with web search billed at $7.00 per 1,000 calls and code execution at $0.12 per hour. Billing is in US dollars, with yen invoicing reserved for enterprise plans. The API is not currently offered in the EU/EEA, the UK or Switzerland while the company completes GDPR work.
Sakana's own use-case posts point at unglamorous volume work: autonomous weekly market-research reports that plan their own search loops and write themselves; customer support that runs from inquiry handling through order-data aggregation; and creative generation, illustrated by a demo where a single prompt about an ocean world led the model to pick motifs, gather reference images by web search, and choreograph a school of roughly a thousand fish.
Why it matters
The interesting part is the supply chain. Japan's flagship AI lab is now productising a Chinese open-weight model as the foundation of a commercial national-language API, competing on price and cultural fit rather than on frontier capability. That is the clearest recent illustration of what open weights from Chinese labs actually enable: a sovereign-language product built in months without a foundation-training run. It also sharpens a policy question for Tokyo, where AI self-sufficiency is a stated goal — Namazu is Japanese where it matters commercially, but its base weights were trained elsewhere.