Today, large language models are trained on vast amounts of text, the majority of which is, by default, in English. This means the model “thinks” well in English and treats everything else as secondary. While languages like Japanese, Arabic, Swahili, and Hindi are formally supported, in practice, the model often struggles with nuances, misses cultural context, or simply provides worse responses than in English.
The Japanese research lab Sakana AI has decided to tackle this very problem. Its team developed an approach called Namazu–a post-training technology that allows taking an already-built large model and “fine-tuning” it for a specific language and culture, without retraining it from scratch.
What Does It Mean to “Adapt” a Model?
Imagine you have a highly educated person who has spent their entire life reading only American books, watching American movies, and working for American companies. They might speak Japanese, but their understanding of how life works, how people should interact, and what is considered polite versus rude would be shaped by a completely different culture. And you'd be able to tell.
Something similar happens with language models. Even if a model technically supports Japanese, it might respond in a way that feels like a “translation from English” rather than natural Japanese speech. It may not know local realities, understand subtext, or take into account the communication norms specific to a country.
Adaptation is precisely about working with this layer: not with the model's foundational knowledge, but with how it communicates, what it deems important, the examples it gives, and how it constructs an answer.
Namazu: Fine-Tune, Don't Retrain
The key idea behind Namazu is not to create a separate model for each language from scratch but to work with existing large, open-source models. To put it simply, you take a ready-made foundation and “tune” it to the required context through post-training.
This is important for several reasons. First, training a large model from scratch is expensive and time-consuming. Second, large models already have a vast knowledge base that you don't want to lose. The goal of post-training is to add desirable qualities without breaking what already works.
Namazu is specifically focused on the largest open-source models–that is, those available to a wide range of developers and researchers, not just major tech companies. This is a crucial point: developers from Japan, India, African nations, or Latin America can use this approach without needing access to closed, proprietary systems.
Why This Matters Right Now
For the past few years, the AI industry has been in a race for scale: who can build a bigger, smarter, more powerful model. This has yielded results–models have indeed become significantly better. But simultaneously, a problem has emerged: the larger the model, the more difficult and expensive it becomes to adapt it for specific needs.
Meanwhile, the need for localization hasn't disappeared. In many countries, AI tools either perform poorly in the local language or aren't available in it at all. This creates an obvious gap: the technology exists, but not everyone can take full advantage of it.
Sakana AI's approach offers a way to bridge this gap. If post-training can indeed effectively adapt large models for different languages and cultures, it paves the way for creating quality local AI tools without having to build everything from scratch every time.
What Questions Remain
For now, Namazu is in its alpha version–an honest acknowledgment that the technology is still developing. How well the approach scales to genuinely diverse languages, especially those that differ significantly from English in structure and cultural context, remains to be tested in practice.
Another open question is how deep this adaptation can be. Post-training is a powerful tool, but it has its limits: if the base model has seen almost no text in a particular language, no amount of fine-tuning will make it truly strong in it.
Nevertheless, the framing of the problem itself seems correct. General-purpose models are great, but the world speaks hundreds of languages, and each has its own logic, its own context, and its own expectations for interacting with a machine. A technology that helps take this into account is not just a technical improvement, but a step toward making AI truly useful to a greater number of people.
