Choosing a Model

The model dropdown after a successful validation

What the job actually asks for

Translation here is a short, structured task: the model receives a JSON array of review titles and bodies and must return the same array, translated, as valid JSON. It rewards instruction-following far more than reasoning power.

That changes the usual calculus. A frontier model is rarely worth the money; a mid-tier model that reliably returns clean JSON is.

Pick by priority

Priority	Pick
Lowest cost per review	A small / flash / mini tier model.
Best idiom and tone	A mid or large tier model.
Data residency	Ollama on your own server.
Fewest malformed replies	A mid-tier model from a major provider.

Two model classes to avoid

Reasoning models are slow and expensive here

A model that "thinks" before answering can take minutes on a 20-review batch and burns tokens on reasoning you never see. If you use one, raise LLM Max Token, and on Ollama raise the request timeout too.

Very small models drift

Sub-2B models often ignore the JSON contract or translate into the wrong language. Every review in a malformed batch silently falls back to its original text — see Fallback Behaviour.

Test before committing

Point a store view at a candidate model, run a small batch, and read the output:

php bin/magento Translate:All -s 2 -b 5

Then check what landed:

SELECT review_id, LEFT(review_message_summary, 60), LEFT(review_message, 80)
FROM   ai_review_translation_list
WHERE  store_id = 2
ORDER  BY entity_id DESC
LIMIT  5;

If the text is unchanged from the original, the model is failing the format — try a larger one before blaming the translation quality.

Next: Switching Providers.