Japanese-developed LLMs for business — ELYZA, Sakana AI, NTT tsuzumi and PLaMo
Japan has its own LLM vendors, and in most English-language coverage they do not appear at all. For a company deploying inside Japan they are worth understanding — not because they beat the frontier models on capability, but because they answer questions the frontier models cannot.
AI Tools Hub Editorial · published · updated
The short version
- For general capability, Claude, ChatGPT and Gemini remain ahead, including in Japanese
- Domestic options matter when data residency inside Japan is a hard requirement
- They also matter for public sector and regulated procurement where a Japanese vendor is preferred or required
- Evaluate them against the requirement, not against GPT-class benchmarks
Why this category exists
Three things drive demand for a domestically developed model in Japan. Data residency — some organisations cannot send data abroad at all. Procurement preference — public sector and some regulated industries weight a domestic vendor heavily. And support reality — an incident at 3am Japan time is handled differently by a vendor with a Tokyo office than by one without.
None of those are capability arguments, and vendors that pitch on capability alone against the frontier labs are making a losing argument. The ones doing well pitch on the three points above.
The main options
| Vendor | Position | Where it fits |
|---|---|---|
| ELYZA | Japanese-tuned models, originating from University of Tokyo research, now under KDDI | Companies wanting a Japanese vendor relationship with reasonable general capability |
| Sakana AI | Tokyo-based research lab with a distinctive approach to model construction | Organisations interested in novel architectures and research partnership |
| NTT tsuzumi | NTT's lightweight Japanese model, pitched on efficiency and on-premise deployment | Regulated industries and public sector where NTT is already a supplier |
| PLaMo | Preferred Networks' model, with a strong engineering reputation behind it | Technically sophisticated teams, including on-premise deployment |
Positions move quickly in this category. Treat the table as a starting point for a conversation with each vendor rather than a current specification sheet.
How to evaluate them honestly
- Test on your own documents, not on public benchmarks — the gap between the two is largest exactly here
- Ask where inference runs, and get it in writing
- Ask about the deployment model: API, VPC, or genuinely on-premise
- Compare total cost including the engineering time an on-premise deployment consumes
- Run the same tasks through Claude or GPT as a control, so you know what you are trading away
When the answer is still a frontier model
If your requirement is general capability in Japanese — summarising, drafting, reasoning over long documents — Claude and ChatGPT are still ahead, and the business plans of both can be contracted in Japan with invoice payment. Choosing a domestic model for those tasks on the assumption that it must be better at Japanese is the most common mistake we see.
Are Japanese models better at Japanese?
Not reliably, in general use. They are trained with more Japanese data proportionally, but the frontier models have scale on their side and handle Japanese well. Test on your own material before assuming either way.
Can we run these on-premise?
Several offer it, which is the strongest argument for the category. Budget for the GPU infrastructure and the people to run it — that cost usually exceeds the licensing difference.
What about public sector procurement?
Domestic vendor preference is real and sometimes decisive. If you are selling into Japanese public sector, this is the category you are competing against, regardless of your benchmark scores.
Related
More articles
Opinion
Why AI rollouts fail in Japanese companies — four structural reasons
For business
Enterprise AI rollout checklist — SOC 2, ISO 27001, DPAs and personal data
Comparison
AI coding tools compared — Cursor, Copilot, Claude Code, Windsurf and Codeium
Ranking
The best AI meeting note tools in 2026 — and whether you need Copilot for it