Sakana Namazu: Kimi K2.6 Retrained for Japanese Business Work

A strong multilingual model is not the same thing as a model that behaves well in a particular country. It can write grammatical Japanese and still pitch the politeness level wrong in a client email, refuse a reasonable question about regional history, or carry the framing of whoever built it. Sakana Namazu is an attempt to fix that layer without training a new model from scratch.

It is Moonshot AI's open Kimi K2.6, post-trained by Sakana AI on in-house data for Japanese and for Japanese business practice. Sakana launched the API on 3 August 2026; OpenRouter's catalogue dates its sakana/sakana-namazu listing to 11 August.

Adaptation, not a new base model

Namazu is the product of a technique Sakana first showed in March 2026: post-training methods that adapt a large open-weight model to a country's culture, values and security requirements. The early alpha series applied it to three different bases, DeepSeek V3.1 Terminus, Llama 3.1 405B and gpt-oss-120B, and Sakana reported that general benchmark performance stayed close to each base while its neutrality and accuracy measures improved. On an in-house set of politically sensitive questions, the DeepSeek base refused 72% and the adapted version reduced that to nearly zero.

The commercial Namazu applies the same approach to Kimi K2.6. That choice brings K2.6's properties with it: a mixture-of-experts model of about one trillion parameters with roughly 32B active per token, vision input, and a 256K context window. OpenRouter lists Namazu with a 262,144-token context, which is that same 256K. If you want the base model's details and trade-offs, the Kimi K2.6 guide covers them.

What the evaluations say

Sakana's results split into two groups. On general capability, AIME26, MMLU-Pro and LiveCodeBench v6, it says Namazu preserves K2.6's performance. On Japanese-specific measures it says Namazu beats the base:

• FairPoliticsQA, a test of neutral answers that do not favour one country's values: 34.10% for Kimi K2.6, 56.30% for Namazu.
• JFBench, Japanese instruction following across one, two, four and eight constraints: improved over the base, with figures shown only in a chart.
• An in-house Japanese translation benchmark: improved, again without numbers in the text.

Treat the FairPoliticsQA result carefully. It is an internal neutrality benchmark, and when Sakana introduced the approach in March it described its in-house measure as whether answers on politics, history and diplomacy involving Japan and other countries present multiple perspectives objectively. That is a reasonable goal, but whether it matches what your users expect is something only your own test set can tell you.

There is one external data point. Aillis, which uses Namazu in a physician-facing evidence search tool, reports scores of 96.8% and 96.4% on Japan's 119th and 120th national medical licensing exams, excluding image questions. That is a single deployment with its own retrieval and citation layer, so it says as much about the system as about the model.

Built-in tools change the cost picture

Namazu ships with web search and code execution as built-in tools, so one request can research, write code and run it. Sakana prices them separately: $7 per 1,000 search calls, with page contents included, and $0.12 per hour of code execution with the session persisted.

Token pricing is $0.95 per million input, $4.00 per million output and $0.15 per million cached input, with no monthly fee, and OpenRouter listed the same token rates at the time of writing. For comparison, OpenRouter listed Kimi K2.6 itself at $0.65 input and $3.41 output at the time of writing, so the base model can be cheaper per token through third-party hosts. Billing is in US dollars; yen figures on Sakana's page are reference prices.

Japanese text also tokenises differently from English, so the same message can cost a different number of tokens on different models. Token costs by language explains why, and it is worth measuring on your own documents before comparing Namazu's bill with an English-centric alternative.

When it is the right pick

Namazu makes most sense when the output itself is Japanese and goes to people: customer support replies, business correspondence in appropriate keigo, internal reports, and research agents that search Japanese sources and summarise them. Sakana's own examples are a weekly market research agent and high-volume support and order-data analysis at a low unit cost.

It makes less sense when the work is language-neutral, such as code, maths or English-only pipelines. There, Sakana's own claim is only that Namazu keeps the base model's performance, so you gain nothing by choosing the adapted version and you take on its access limits.

Those limits are real. Namazu is not yet available in the EU, EEA, the UK or Switzerland. Inputs may be used for training by default, which you can turn off in the console. Sakana does not guarantee that processing stays within Japan; region selection, yen invoicing and zero data retention are offered as enterprise discussions. Sakana's pages describe Kimi K2.6 as an open model but do not say that Namazu's own weights are released, so plan on API access only.

A short evaluation plan

1. Collect 30 real Japanese tasks: a few emails to clients, a few support replies, a few summaries of Japanese documents.
2. Run each on Namazu and on Kimi K2.6 with the same prompt.
3. Have a native speaker rate them blind for register and naturalness, not just correctness.
4. Add five questions on regional history or politics that matter to your users and compare how each model frames them.

OpenRouter lists a single variant with no free tier. If the native-speaker ratings favour Namazu and the access restrictions do not rule you out, it is a cheap upgrade over the base model for Japanese-facing work.