Key point: Namazu is a single model specialized for Japanese, while Fugu Max is an orchestrator that coordinates multiple models. The memory feature is available on both.

1. Introduction

In the model selection menu of Sakana Chat, you'll see "Namazu" and "Fugu Max" listed side by side. The names alone don't make the difference obvious, so you may be unsure which one to pick.

Sakana AI is developing a mobile app, and on November 9, 2026, it announced a call for beta testers on its official X account. In this article, we break down the differences between the two models available on the web version, based on publicly available information.

2. Main Features (Overview of the Two Models)

  • Namazu: A single LLM specialized for Japanese and Japanese business practices. It is based on Moonshot AI's open model "Kimi K2.6" and has been further trained (post-trained) by Sakana AI on proprietary in-house data. It has been available since Sakana Chat launched in March 2026.
  • Fugu Max: An orchestrator model that coordinates multiple models. It is itself a language model, and it dynamically assembles agent configurations and processing steps based on the user's question, then executes them. From the API, it can be called like a single model.

The main updates to Sakana Chat since its public release are as follows:

  • March 2026: Launched alongside the alpha version of Namazu
  • August 13, 2026: Fugu was added, making model selection possible. Code execution and image/document attachments were supported
  • September 17, 2026: Fugu was overhauled into Fugu Max and made available to all users. The memory feature, usable with both Namazu and Fugu Max, was also added

Namazu also became available via API on August 3, 2026. The API is OpenAI-compatible and includes web search and code execution as built-in tools by default.

3. How It Works

3-1. Namazu's Post-Training

Namazu has undergone additional training on proprietary in-house data on top of its base model. According to official information, the post-training covers the following:

  • Honorifics and business practices: Adapting to natural Japanese for document and email processing
  • Reducing excessive refusals: Adjusting responses so the model doesn't decline too often
  • Suppressing output bias: Adjusting to reduce bias toward the values of specific countries

According to Sakana AI's own evaluations, Namazu maintained the base model's reasoning ability (AIME26, MMLU-Pro, LiveCodeBench v6) while outperforming the base model on Japanese instruction following (JFBench), Japanese-English translation, and response neutrality. The FairPoliticsQA score for response neutrality reportedly rose from 34.10% to 56.30%.

All of these are the company's own evaluations and claims, not independently verified by third parties. The translation and FairPoliticsQA figures are in-house benchmarks.

3-2. Fugu Max's Orchestration

Rather than having humans define roles in advance, Fugu Max assembles specialist agents from a model pool according to the task and assigns the processing work. Models in the pool can be swapped out, and specific models can also be excluded.

The technical foundations are the ICLR 2026 papers "TRINITY" and "Conductor." In TRINITY, an evolutionary lightweight coordinator assigns each model the role of Thinker, Worker, or Verifier. In Conductor, the natural-language coordination strategies themselves are learned through reinforcement learning.

AB-MCTS (arXiv:2503.04412) is also relevant prior research. However, it has not been officially stated that it is used inside Fugu.

The Fugu family consists of the following four models:

  • Fugu: A balanced model for performance and speed
  • Fugu Ultra: A model focused on performance
  • Fugu Max: A model focused on cost performance, with the largest model pool
  • Fugu Cyber: A model specialized for security

The one you can select in Sakana Chat is Fugu Max.

4. Usage (Rough Guidelines for Choosing)

  • When to choose Namazu: Repeated workflows that produce natural Japanese outputs from Japanese instructions, such as processing documents and emails that require honorifics and knowledge of business practices. It is also suitable when response neutrality matters.
  • When to choose Fugu Max: When you want to combine the strengths of multiple models, or when you want agents to be assembled according to the task.

The memory feature, which carries conversation context forward, is available on both Namazu and Fugu Max.

API pricing is as follows (per 1 million tokens):

Model Input Output Notes Namazu $0.95 $4.00 From December 1, 2026 Input $1.90 / Output $6.00 scheduled Fugu Max $2 $6 

5. Applications (Official Use Cases)

Official use cases for Namazu include automated generation of market research reports and customer support automation. The intended use is repeated workflows that take in Japanese instructions and return natural Japanese outputs.

Fugu Max is set up to distribute processing to models according to the task. A use case would be tackling problems that are hard for a single model by combining multiple models.

6. How This Differs from the Past (Contrast with the Single Massive Model Approach)

Sakana AI promotes three pillars: "pre-training, post-training, and orchestration." Namazu is positioned as the post-training product, and Fugu as the orchestration product.

In the official FAQ, Fugu is described as a service that orchestrates multiple models, while Namazu is described as a single proprietary model that has been additionally trained in-house.

This is a different approach from the path of pre-training a single massive model in-house. A difference in approach alone does not determine which is better in terms of performance.

7. Caveats

  • Benchmarks: All numbers are the company's own evaluations and claims, not verified by third parties. They do not establish superiority or inferiority over competing products.
  • Technical report: The benchmarks in arXiv:2606.21228 are for Fugu and Fugu Ultra; they are not figures for Fugu Max.
  • Model pool: The setup allows models to be swapped out. The official explanation is that this protects users from vendor lock-in, API discontinuation, and geopolitical uncertainty.
  • AB-MCTS: It has not been officially stated that it is used inside Fugu.
  • Mobile app: It is in development, and a call for beta testers has been officially announced.

8. Summary

Namazu is a single model post-trained for Japanese and Japanese business practices. Fugu Max is an orchestrator that dynamically assembles multiple models.

For Japanese document and email processing, Namazu is the choice; for problems that require coordination among multiple models, it's worth trying Fugu Max. The memory feature is available on both.

9. The Author's Take

Having used both models, Namazu felt designed for users who prioritize response speed. With Fugu Max, I got a strong sense of accuracy in answers to heavy tasks and complex consultations. This is purely the author's subjective experience.

Going forward, I think the question will be how Sakana AI differentiates itself between cutting-edge models like ChatGPT and Claude and lightweight models coming out of China. Optimization for Japanese and the evolution of the mechanism for bundling models together look set to be key.

10. References