Tokyo-based AI startup Sakana AI has developed a specialized post-training technique designed to adapt world-class open-weight foundation models to Japan's unique culture, values, and security requirements. Currently, pre-training large language models (LLMs) requires such massive computational resources that cutting-edge development is largely concentrated among US and Chinese tech giants. However, as the movement toward open-weight models—where trained internal data is made public—gathers momentum, the key challenge for domestic industries has shifted to refining these global assets into Japan-specific solutions.

The newly announced Namazu (Alpha) prototype series serves as a powerful answer to this challenge. Built upon world-leading models like DeepSeek-V3.1 and Llama-3.1-405B, these versions have undergone Sakana AI’s proprietary post-training process. Remarkably, this adaptation to the Japanese context is achieved without compromising the models' original high-level reasoning, coding expertise, or mathematical capabilities. Benchmark results confirm that Namazu maintains foundational intelligence equivalent to its original base models.

Furthermore, the quality of responses has improved dramatically for sensitive themes such as politics, history, and diplomacy, where regional ideologies often influence AI outputs. While conventional foreign-made models frequently exhibit excessive refusals or biases when faced with specific local topics, Namazu significantly enhances neutrality and factual accuracy. For instance, where the original DeepSeek-based model refused roughly 70% of sensitive questions, the Namazu version reduced that refusal rate to nearly zero, offering multifaceted perspectives grounded in objective facts.

In addition to the models, Sakana AI has launched Sakana Chat, a public chat service with web search functionality powered by Namazu. This allows users to experience a Japanese-specification AI that can reference real-time information during dialogue. Moving forward, the company plans to integrate agent technologies and advanced control systems for multiple models, aiming to deliver next-generation solutions that transcend simple chat interfaces. This initiative represents a milestone for Sovereign AI, creating a model for importing global intelligence while preserving national identity.

Tokyo-based AI startup Sakana AI has developed a specialized post-training technique designed to adapt world-class open-weight foundation models to Japan's unique culture, values, and security requirements. Currently, pre-training large language models (LLMs) requires such massive computational resources that cutting-edge development is largely concentrated among US and Chinese tech giants. However, as the movement toward open-weight models—where trained internal data is made public—gathers momentum, the key challenge for domestic industries has shifted to refining these global assets into Japan-specific solutions.

The newly announced Namazu (Alpha) prototype series serves as a powerful answer to this challenge. Built upon world-leading models like DeepSeek-V3.1 and Llama-3.1-405B, these versions have undergone Sakana AI’s proprietary post-training process. Remarkably, this adaptation to the Japanese context is achieved without compromising the models' original high-level reasoning, coding expertise, or mathematical capabilities. Benchmark results confirm that Namazu maintains foundational intelligence equivalent to its original base models.

Furthermore, the quality of responses has improved dramatically for sensitive themes such as politics, history, and diplomacy, where regional ideologies often influence AI outputs. While conventional foreign-made models frequently exhibit excessive refusals or biases when faced with specific local topics, Namazu significantly enhances neutrality and factual accuracy. For instance, where the original DeepSeek-based model refused roughly 70% of sensitive questions, the Namazu version reduced that refusal rate to nearly zero, offering multifaceted perspectives grounded in objective facts.

In addition to the models, Sakana AI has launched Sakana Chat, a public chat service with web search functionality powered by Namazu. This allows users to experience a Japanese-specification AI that can reference real-time information during dialogue. Moving forward, the company plans to integrate agent technologies and advanced control systems for multiple models, aiming to deliver next-generation solutions that transcend simple chat interfaces. This initiative represents a milestone for Sovereign AI, creating a model for importing global intelligence while preserving national identity.