Sovereign LLM selection: five criteria that decide it

Why starting from accuracy fails

As of 2026 the major international models handle Japanese perfectly well. Domestic models do sometimes top Japanese benchmarks, but the situations where that margin changes a business outcome are rare. If there is a reason to consider a domestic model, it is not accuracy — it is where the data can sit and what contracts are available.

Which also means that if nothing in your requirements demands a domestic vendor, restricting the shortlist to domestic models has no purpose. Start by writing down why you are considering one.

Criterion 1: delivery model

Domestic models are delivered in three broad ways.

Commercial API: requests go to the vendor's cloud. Fast to adopt; data leaves your estate
Open weights: the weights are published and you run them yourself. GPUs required
Negotiated enterprise supply: quoted and configured to requirements

ELYZA sells primarily to enterprises, with terms settled through enquiry. Some vendors also publish models for research and evaluation, so the same organisation has different front doors by use case. Approaching them without having decided which model of delivery you want produces a conversation that goes nowhere.

Criterion 2: where processing happens

This is usually the real reason for the whole exercise. Two questions: where inference requests are processed, and whether that data is retained or used for training.

An API from a domestic vendor can still run on infrastructure hosted abroad. 'Domestic model' and 'domestic processing' are not the same claim. Get the processing region in writing before signing.

Criterion 3: licence

For open-weight models the licence decides whether and how you may use it commercially. Permissive licences such as MIT and Apache 2.0 sit alongside bespoke community licences that can restrict scale or use case.

Is commercial use permitted?
Who owns the output, and who carries liability for it?
May the model be modified and redistributed?
Does a base model's licence flow through?

The fourth is the one that gets missed. Several Japanese-adapted models are built on international open models, in which case the base model's terms apply as well. Check what the base is, not only the licence label on the page you are reading.

Criterion 4: the GPUs you can actually get

Running a model yourself ties required GPU memory to parameter count. A 100B-class model needs a multi-GPU configuration, which raises both procurement and operational difficulty.

Models deliberately sized to run on a single GPU exist precisely to make on-premise deployment realistic. NTT's tsuzumi is positioned around single-GPU operation.

Decide which GPUs you can obtain first, then choose from the models that fit. Choosing the model first and then sourcing hardware means procurement lead times break the plan.

Criterion 5: commercial continuity

A model being published does not mean it will continue to be supplied. Foundation model development is capital-intensive and strategies change within a year or two. If it is going into a business process, look for evidence that the vendor intends to keep selling it commercially.

Public-sector procurement outcomes are a useful signal. Japan's Digital Agency ran a selection process for domestic LLMs in 2026 under which several models were reported as adopted. Reporting of the list and its conditions varies between secondary sources, so use the official publication as the primary reference if it informs your decision.

The order to work in

Order	Decide	What goes wrong otherwise
1	Why a domestic model at all — data requirement or procurement requirement	The comparison collapses into accuracy and never concludes
2	Delivery model: API, self-run, or negotiated	Vendor conversations stall
3	The GPU ceiling you can procure	You select a model you cannot run
4	Licence and commercial terms	Legal stops it the week before launch
5	Processing region and training use	IT review sends it back

Skipping step one is the most common failure. Without the reason in words, no model can be chosen.

Frequently asked questions

Is a domestic model better at Japanese?

It depends on the task. Some top Japanese benchmarks, but as of 2026 the major international models handle Japanese well. Accuracy alone is a weak basis for the choice; data and procurement requirements are the real ones.

Self-hosting or API — which is cheaper?

It inverts with volume. At low volume the API is dramatically cheaper; at sustained high volume self-hosting can win. But self-hosting adds GPU procurement and operational staffing, so the break-even sits at a high volume.

How far can this article be trusted?

It aggregates public information, and model versions, pricing and licences change. Use each vendor's own materials as the primary source; this piece supplies the selection framework.

Choosing a sovereign LLM — five criteria, and why accuracy is not one