Sakana's Namazu, a Kimi K2.6-based Japanese LLM, lands on OpenRouter
The Moonshot-derived model is now an OpenAI-compatible endpoint with built-in web search and code execution — as Alibaba's Qwen3.8 open weights loom this week.
Copy markdown
A Japanese-tuned fork of Moonshot's Kimi K2.6
Sakana AI built Namazu on Moonshot's open Kimi K2.6, then fine-tuned it in-house for Japanese language, honorifics, and business context. If you ship for Japanese users — or want a Kimi-class model specialized rather than generic — it's a genuinely new option, though Sakana hasn't published its training method or benchmark tables.
Now one OpenRouter call away, tools built in
Namazu has surfaced on OpenRouter next to its OpenAI-compatible API, so you can reach it without a Sakana account and drop it into existing SDK code by swapping the base URL. It exposes web search and code execution as native tools, so there's no extra wiring. Pricing and context window still aren't officially disclosed, so confirm both before you commit a project to it.
Watch: Qwen3.8 open weights are due any day
Alibaba's Qwen3.8-Max API (2.4T parameters, 95B active, 1M-token context) has been live since Aug 4, and Alibaba promised open weights for both Max and a 27B companion 'the week of Aug 10.' Nothing has hit Hugging Face yet and the license is unconfirmed — so if you're planning a local deploy, watch the Qwen org page and read the LICENSE file before you pull anything.