The "Japanese language problem of foreign LLMs" that Namazu solves
On March 24, 2026, Sakana AI unveiled "Namazu (alpha version)," a Japanese-specialized LLM series, along with "Sakana Chat," an AI chat service powered by it.
Using foreign-made large language models as-is in Japanese leads to several issues:
- Response refusals: Excessive filtering on questions regarding politics, history, and diplomacy
- Cultural bias: Shallow understanding of Japan-specific contexts (legal systems, business customs, etc.)
- Unnatural Japanese: Literal translations and phrasing that native Japanese speakers would not use
Namazu resolves these issues through proprietary "post-training" technology.
Namazu's technical approach — What is post-training?
At the core of Namazu is post-training technology that adapts existing high-performance models to Japanese specifications. Instead of building models from scratch, it maintains the underlying performance of foundation models like DeepSeek and Llama while tuning their behavior for use in Japan.
Model lineup
| Model | Base model | Parameter scale |
|---|---|---|
| Namazu-DeepSeek-V3.1-Terminus | DeepSeek-V3.1 | 685 billion |
| Llama-3.1-Namazu-405B | Llama 3.1 | 405 billion |
| Namazu-gpt-oss-120B | GPT-oss | 120 billion |
Dramatic reduction in refusal rates
The most striking achievement is the improvement in response refusal rates.
While the base model DeepSeek-V3.1-Terminus refused to answer 72% of questions regarding politics, history, and diplomacy, Namazu improved this to nearly 0%. Its reasoning, knowledge, and coding performance remain roughly equivalent to the base model.
Benchmark evaluation
Namazu is evaluated across three dimensions:
| Evaluation axis | Result |
|---|---|
| Core capabilities (MMLU-Redux, GPQA, etc.) | Reasoning, knowledge, and coding performance roughly equivalent to the base model |
| Neutrality and factual accuracy | All three models showed improvement over their base models in independent evaluations |
| Japanese proficiency (Nejumi, Swallow, etc.) | Japanese performance on par with models of comparable scale |
How to use Sakana Chat
Sakana Chat is available free of charge with no account registration required.
- Access chat.sakana.ai
- Select a model (default is Namazu-DeepSeek-V3.1-Terminus)
- Enter your question and submit
Web search is integrated, allowing answers that include real-time information. However, access is restricted to users within Japan.
How to use it alongside Claude and GPT
Namazu is not an all-purpose tool; it demonstrates its strengths in specific use cases.
| Use case | Recommended model | Rationale |
|---|---|---|
| Japanese politics, history, and diplomacy | Namazu | Provides neutral answers without response refusals |
| Natural writing in Japanese | Namazu / Claude | Phrasing grounded in cultural context |
| Advanced coding | Claude / GPT | Development tool integrations and proven track record |
| Data analysis and reasoning | Claude / GPT | Advantage in benchmark scores |
| Processing confidential data | Namazu | As an option that stays entirely within Japan |
| Quick and free experimentation | Namazu（Sakana Chat） | No registration required and free |
Potential as sovereign AI
The true value of Namazu lies not only in its "Japanese proficiency," but also in the context of data sovereignty (sovereign AI). For enterprises wanting to run AI on Japanese infrastructure without transmitting data to foreign services, Namazu can be a compelling choice.
Current limitations
Given that this is an alpha release, keep the following points in mind:
- Detailed benchmark scores and technical reports are scheduled for release at a later date
- API is not yet available (web access via Sakana Chat only)
- Model weights are published on Hugging Face, but commercial licensing depends on the base model
Conclusion
Sakana AI's Namazu takes the practical approach of "adapting the strongest foreign models to Japanese specifications" rather than "building a Japanese LLM from scratch."
- Achieved refusal rates dropping from 72% to nearly 0% via post-training technology
- Sakana Chat is free and registration-free, with web search already integrated
- Practically best used as a complementary option rather than a replacement for Claude or GPT
The formal release of technical reports and the launch of API offerings will be the next key milestones to watch. Expanding options for Japanese-capable AI is a welcome development for users.
Related articles
- Practical Guide to Running Gemma 4 Locally — Another notable open model
- ChatGPT Business Playbook: 10 Prompts for Practical SME Work — Practical examples of LLMs in business
- Cursor 3 vs. Claude Code: An In-Depth Comparison — How to choose AI development tools
