Sakana AI Releases Namazu, a Japanese-Specialized Reasoning Model Built on Kimi K2.6

Sakana AI has released Namazu, a reasoning model specialized for Japanese language and business use cases, built on top of Moonshot AI's Kimi K2.6 architecture with additional training data.

What Namazu Does

According to Sakana AI, Namazu is designed for Japanese instruction following, business writing, mathematics, coding, research, and multi-step agent workflows. The model inherits its base capabilities from Kimi K2.6 and adds targeted fine-tuning for Japanese linguistic and business-context tasks — a niche that general-purpose frontier models often handle less precisely than dedicated regional models.

Pricing and Specs

Namazu is priced at $0.95 per 1 million input tokens and $4 per 1 million output tokens. The model supports a context window of 262,000 tokens, putting it in line with other long-context reasoning models currently available through OpenRouter.

No published benchmark scores accompanied the release. Sakana AI has not disclosed a parameter count or training cutoff date for Namazu.

Data Handling Notice

Sakana AI states that it may use user inputs for model training and improvement by default, with an opt-out available through the Sakana Console. The company also notes that processing of requests is not guaranteed to occur entirely within Japan, a detail relevant to enterprises with data residency requirements — a common consideration for Japanese business customers evaluating AI vendors.

What This Means

Namazu is a fine-tune rather than a from-scratch model, built on Moonshot AI's Kimi K2.6 base. This approach lets Sakana AI compete in the Japanese-language and business-context niche without the cost of training a foundation model from zero. The pricing — $0.95/$4 per 1M tokens — positions Namazu as a mid-tier option, more expensive than budget open models but well below flagship frontier pricing.

The lack of published benchmarks makes it difficult to independently verify Sakana AI's claims about Namazu's performance on Japanese instruction following or coding tasks. Buyers evaluating the model for production use, particularly in regulated industries, should also weigh the data residency caveat: Sakana AI explicitly does not guarantee in-country processing, which could matter for Japanese enterprises with strict compliance requirements. As with any derivative model, its ceiling is ultimately bound by the capabilities of its Kimi K2.6 base plus whatever additional fine-tuning Sakana AI applied.