Sakana Namazu

Sakana Namazu

Quick answer: Sakana Namazu is Sakana AI's Japanese-specialized LLM API, released August 3, 2026, built by fine-tuning Moonshot AI's open-weight Kimi K2.6 on in-house Japanese business-context data. It preserves (and slightly improves on) Kimi K2.6's reasoning and coding scores — 96.67% AIME26, 90.33% MMLU-Pro, 90.33% LiveCodeBench v6 — while substantially improving Japanese-specific performance, most notably FairPoliticsQA (34.10% → 56.30%). It ships with built-in web search and code execution tools behind an OpenAI-compatible API, priced at $0.95 / $4.00 per 1M input/output tokens.

At a Glance

Where Sakana Namazu leads

• 96.67% AIME26 and 90.33% MMLU-Pro — matches or slightly exceeds its Kimi K2.6 base model on core reasoning/knowledge benchmarks
• 56.30% FairPoliticsQA — up from the base model's 34.10%, its largest single improvement, reflecting explicit tuning for politically-neutral, non-US-centric answers
• 37.40% JFBench and 52.20% internal Japanese-English translation — outperforms Kimi K2.6 on every Japanese-specific evaluation tested
• Built-in web search and code execution tools, usable via a drop-in OpenAI-compatible API (just change base_url and the model name)
• Low, usage-based pricing with no monthly fee — $0.95/$4.00 per 1M input/output tokens, $0.15 per 1M cached input tokens

Where it lags

• API-only, proprietary — no open weights, unlike its Apache 2.0-licensed Kimi K2.6 base
• Not currently available in the EU, EEA, UK, or Switzerland (regulatory compliance pending)
• No guarantee of in-Japan-only data processing at this time

Best for: Japanese-language business workflows — customer support automation, market research agents, and document/email processing — where natural, honorific-appropriate Japanese and low per-token cost matter more than access to open weights.

What Sakana Namazu Is

Sakana Namazu is Sakana AI's production API version of the model previously powering Sakana Chat, rebuilt on top of Moonshot AI's open-weight Kimi K2.6. Rather than training a model from scratch, Sakana AI fine-tuned Kimi K2.6 on proprietary in-house data to adapt it to Japanese language conventions and Japanese business context — natural keigo (honorific) usage, common business-document and email formats, and Japan-specific factual and cultural knowledge — while also tuning down unnecessary topic refusals and reducing bias in its outputs toward any single country's values.

The model retains the agentic capability and reasoning strength of its Kimi K2.6 base (matching it on AIME26, MMLU-Pro, and LiveCodeBench v6, per Sakana's own published comparison) while measurably improving on every Japanese-specific evaluation the company tested: JFBench (Japanese instruction-following), an internal Japanese-English translation benchmark accounting for proper nouns and honorifics, and FairPoliticsQA (neutrality of answers across national viewpoints), where it posted the largest gain of the group.

Sakana Namazu is offered exclusively as a hosted, OpenAI-compatible API (distinct from Sakana Fugu, which orchestrates multiple frontier models rather than running as a single proprietary model) with built-in web search and code-execution tools, aimed at Japanese enterprises that want production-ready Japanese-language quality without either the cost of frontier proprietary APIs or the operational burden of self-hosting an open model.

Specifications

Pricing

No monthly subscription fee — usage-based billing only, in USD (with a JPY reference conversion shown in the Sakana console at ¥160/$1).

Public Benchmark Scores

Scores are self-reported by Sakana AI in the official Sakana Namazu launch announcement, directly comparing Namazu against its Kimi K2.6 base model.

Notable In-House / Unlisted Benchmark Results

Sakana AI also reports results on two evaluations explicitly described as internally built (not independently published or verifiable), so they aren't tracked as standalone Benchgen leaderboard entries:

• Japanese-English Translation (internal) — 52.20%, up from the Kimi K2.6 base's 47.80%. Tests translation quality accounting for Japan-specific proper nouns and honorific language.
• FairPoliticsQA (internal) — 56.30%, up from the base model's 34.10% — the largest improvement Sakana reported, measuring whether answers stay neutral rather than leaning toward any one country's values.

Sakana Namazu vs Kimi K2.6 (Base Model)

Sakana Namazu preserves — and in most cases very slightly exceeds — Kimi K2.6's reasoning and coding performance, while delivering a clear, consistent uplift across every Japanese-specific benchmark tested. The FairPoliticsQA gain (+22.2 points) is by far the largest, reflecting deliberate tuning for balanced, Japan-aware responses rather than a US-centric default.

FAQ

Is Sakana Namazu open source? No — Namazu is offered only as a hosted, proprietary API. Its base model, Kimi K2.6, is separately available as an open-weight (Apache 2.0) model from Moonshot AI, but Sakana's fine-tuned Namazu checkpoint itself is not released.

What is Sakana Namazu built on? Kimi K2.6, the open-weight mixture-of-experts model from Moonshot AI, fine-tuned by Sakana AI on proprietary Japanese business-context data.

How much does Sakana Namazu cost? $0.95 per 1M input tokens and $4.00 per 1M output tokens, with cached input at $0.15 per 1M tokens. Web search costs $7.00 per 1,000 calls; code execution costs $0.12 per hour of session time. No monthly fee.

How is Sakana Namazu different from Sakana Fugu? Sakana Fugu orchestrates multiple frontier models together to boost performance; Sakana Namazu is a single, independently fine-tuned proprietary model served as its own API.

What is Sakana Namazu best used for? Japanese-language business workflows — customer support automation, automated market-research agents, and document/email processing — where natural, keigo-appropriate Japanese responses and low token cost matter more than open weights or non-Japanese use cases.

Benchmark scores sourced from Sakana AI's official Sakana Namazu launch announcement (sakana.ai/namazu-api, Aug 3, 2026) and the Sakana Namazu product page (sakana.ai/namazu).