Developed post-training technology to adapt largest-scale open foundation models to national specifications.

2026-03-24 · Source: Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Intermediate, medium

Summary

Sakana AI has developed and released the "Namazu" series, an alpha version of prototype models that adapt existing frontier open-weight foundation models to Japanese cultural values and security requirements through post-training technology. Concurrently, the company launched "Sakana Chat," a chat service featuring the Namazu models, which include Namazu-DeepSeek-V3.1-Terminus, Llama-3.1-Namazu-405B, and Namazu-gpt-oss-120B. The Namazu series maintains performance comparable to its base models across major benchmarks like AIME’25 and MMLU-Redux, while significantly improving neutrality and factual accuracy on politically sensitive topics related to Japan. It also reduces refusal rates on such questions from 72% to nearly 0% for models like DeepSeek-V3.1-Terminus. Sakana Chat integrates a Web search function to provide real-time information.

Key takeaway

For NLP Engineers developing LLM applications for specific national or cultural markets, Sakana AI's Namazu series demonstrates that post-training can effectively mitigate inherent biases and censorship tendencies of global foundation models. You should consider implementing similar post-training methodologies to ensure your models provide neutral, accurate, and culturally appropriate responses, especially for sensitive topics, without compromising core performance. This approach can significantly enhance user trust and regulatory compliance in target regions.

Key insights

Post-training adapts frontier LLMs to specific cultural and security contexts while preserving performance.

Principles

• Open-source LLMs require post-training for regional bias correction.
• Cultural adaptation can reduce model refusal rates on sensitive topics.

Method

Proprietary post-training technology uses unique datasets for bias correction within specific cultural contexts, applied to high-performance open-weight foundation models.

In practice

• Use post-training to align LLMs with local cultural norms.
• Integrate Web search for real-time information in chat services.

Topics

• Post-Training Technology
• Foundation Model Adaptation
• LLM Bias Correction
• Country-Specific AI
• Sakana Chat

Best for: NLP Engineer, AI Engineer, Machine Learning Engineer, AI Product Manager

Related on AIssential

• To Train or Not to Train — Post-training specialized models on open-weights bases offers AI application companies economic, latency, and differentiation advantages, despite rapid base model evolution.
• Modeling LLM Unlearning as an Asymmetric Two-Task Learning Problem — LLM unlearning can be reframed as an asymmetric two-task problem prioritizing retention over forgetting.
• Bridging Cultural Gaps in Automated Translation of Brazilian Expressions: A Study on Cultural Adaptation — Automated translation systems struggle with cultural adaptation, even with advanced LLMs, requiring human oversight.
• Frontier post-training recipe review with Finbarr Timbers — Frontier LLM post-training now demands complex, multi-stage, multi-teacher distillation for advanced reasoning and domain specialization.
• ImportAI 449: LLMs training other LLMs; 72B distributed training run; computer vision is harder than generative text — LLMs can partially refine other LLMs, while post-training significantly boosts AI capabilities.
• Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning — Larger LLMs demonstrate superior efficiency and performance in RL post-training for mathematical reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Blog.