DeepL translation models

Next-generation models and release history

The next-generation line was introduced in July 2024. The release describes training that combines DeepL's linguistic data with feedback from language specialists. It initially highlighted Japanese, German and Simplified Chinese translation to and from English. The current article reports more than 100 supported languages.DeepL next-generation language model

DeepL's May 2026 platform article records a new model released that March. It describes translation customization through glossaries, style rules and translation memories. These are controls around the service; the article does not supply a public checkpoint ID for the March release.DeepL's March 2026 model update and translation platform

What the API selector means

The current API defaults to latency_optimized. quality_optimized requests the highest translation quality. prefer_quality_optimized is a legacy value with the same current behavior as quality_optimized. All supported language pairs accept these values, and all source and target languages have next-generation support as of December 2025.DeepL: Understanding Model Types

DeepL may serve the same model for both goals on some requests. A request that supplies model_type receives model_type_used in the response. The document endpoint accepts this parameter but ignores it. A latency choice therefore does not guarantee access to a particular older classic checkpoint.DeepL: Understanding Model Types

Text and document workflows

Use the text endpoint for strings such as messages and interface text. Its request array preserves result order, but each string is translated independently. Strings in the same array do not share context. A full request must remain within the byte limit; this is not a published model token window.DeepL text translation API reference

Use the document endpoint for complete files. It uploads a document, reports translation status and returns the translated file. A document key is needed for status and download requests. The result is removed after download, so an integration must save the returned file when it receives it.DeepL document translation quickstart

DeepL's reconstruction research explains how page images can be processed for text and layout information, then rebuilt after translation. This can help with scans and complex pages. It does not mean that the text API accepts arbitrary images or provides a general visual question-answering model.DeepL AI Labs: document reconstruction with VLMs

Context and terminology controls

Supply surrounding source content through context when a short phrase is ambiguous. The guide distinguishes this from instructions. Use glossaries for term mappings, and style rules or custom instructions for output requirements. A context paragraph is not a reliable substitute for a system prompt.DeepL context parameter guide

The text reference supports up to ten custom instructions, each with at most 300 characters. Style profiles can package several customization settings under one ID. The translation overview confirms that customization also applies to document workflows.DeepL text translation API referenceDeepL Translate API overview

Practical assessment

Practical guidance. Start a translation review with complete sentences, short interface labels and terms that have a specific business meaning. Compare omission errors, names, numbers and formatting separately. A fluent sentence can still change a condition or omit a qualifier. For labels, include nearby source content; for fixed terminology, define the required translation explicitly.DeepL context parameter guideDeepL Translate API overview

Practical guidance. Record settings and the review date when comparing results. DeepL can change the model behind a selection goal. Keep terminology controls consistent, and assess results for the exact language pair your application needs.DeepL: Understanding Model Types