Models and fallbacks

Select models for language pairs and understand which candidate a request tries when another cannot complete.

A localization engine's model configuration answers two questions: which model should translate a language pair, and what should be tried if it cannot complete the request? Fallbacks address execution failures. They do not generate several translations and choose the one with the highest quality score.

The web app's creation form defaults to a recommended primary model and fallback for all locale pairs. You can keep that broad coverage or choose a more specific configuration during creation or later.

Choose the model and its coverage #

Use the supported model catalog in the configuration interface or account MCP for current identifiers and context limits. Availability changes; a copied list can become stale. With a Mistral AI key, the configuration interface also offers chat models your key serves beyond the catalog; see models beyond the catalog.

An exact model identifier names a particular selection. A moving alias can resolve differently later. Pinning makes the choice easier to explain, but it does not make output deterministic or establish suitability for your content.

Configure the source and target locale pair. Use wildcard coverage for a general default, and specific locales when the pair needs a different candidate. Confirm that important pairs have applicable configuration; retrying cannot repair a missing model configuration.

Read the attempt order #

The localization engine orders candidates by:

  1. Target locale match tier.
  2. Source locale match tier.
  3. Target locale specificity.
  4. Source locale specificity.
  5. Configured rank.

A rank of one does not override a better locale match. For example, a specific English-to-de-CH configuration can precede a broad wildcard configuration even if the latter was entered first. Locale resolution explains which locales apply to a request, how a bare language stands for its default region, and how match tiers rank them.

Attempts have execution deadlines. A failed or timed-out candidate can give way to the next, while exhausting the chain fails the request. A fallback reduces dependence on one model; it cannot supply coverage that is absent or guarantee success when every candidate fails.

Compare quality separately from reliability #

Use request evidence to identify the model actually used and whether a fallback occurred. A provider refusal, a configuration error, and a completed but unsuitable translation need different repairs.

To run a provider's models on your organization's own key, see BYOK. A Lingo.dev API key and an external provider credential authorize different services; selecting a model does not create a credential connection.

The translation API reference covers JSON content and context fields. Every job requires an engineId. Confirm that the selected engine is enabled and has model coverage for each requested locale pair. Inspect the job’s model attempts to establish which model actually handled each batch.