DeepL

What DeepL is and who develops it

DeepL is a family of artificial intelligence services for working across languages. Its best-known product, DeepL Translator, converts existing text into another language; DeepL Write revises written expression and DeepL Voice handles conversations. The company behind them is DeepL SE, based in Cologne, Germany, and led by its founder Jaroslaw, or Jarek, Kutylowski.

Three decisions should be distinguished: which content needs translating, the format required for delivery and the review its intended use demands. Translating an email to understand it, localising a catalogue and publishing a work instruction have different costs of error, even when they use the same language engine.

Reference documentation: DeepL: About us · DeepL Publisher · DeepL: About Translator.

Origins and development

The project emerged from Linguee, whose work with bilingual texts preceded the translator. DeepL opened to the public in August 2017; DeepL Pro followed in March 2018 with subscriptions and professional integration. Russian and Portuguese joined the initial seven languages in December 2018.

• 2023: Write’s public beta was introduced on 17 January, initially in English and German. New investors announced that month included IVP, Bessemer, Atomico and WiL.
• 2024: Write Pro appeared in April. In May, the company announced a $300 million investment at a $2 billion valuation. In July, it introduced its next generation of translation using a specialised LLM, and in November it launched Voice.
• 2025: the November expansion added more than 70 languages, taking the translator beyond 100. The same launch introduced Customization Hub and DeepL Agent.
• 2026: speech-to-speech translation was announced in April, followed in September by connections between DeepL and assistants through MCP. The platform thus added further ways of working alongside the web translation box.

These dates describe specific announcements. A feature’s availability in an account depends on the product, offering and rollout; commercial history does not replace access documentation.

Reference documentation: Slator: Inside DeepL · DeepL: First language expansion · DeepL launches DeepL Write · DeepL: New investor support · DeepL announces $300 million investment · DeepL: Next-generation language model · DeepL launches Voice · DeepL: New tools and DeepL Agent · DeepL: How we launched over 70 new languages · DeepL: Voice-to-voice translation · DeepL MCP in Microsoft Copilot, ChatGPT and Claude.

How translation works

Current documentation describes language-model infrastructure specialised in translation. The system estimates how to express content in the target language using context; it does not simply consult a dictionary and replace each word. Two sentences containing the same word can therefore produce different translations.

DeepL explains that it combines proprietary architectures, multilingual data selection and training methods. It does not publish enough detail to reproduce the complete service. Historical explanations of neural networks should be read with their dates, without assuming that they describe the model processing a request today.

The infrastructure has also evolved: computing in Iceland was followed by collaboration with NVIDIA and EcoDataCenter. The company documents DGX GB200 systems and liquid cooling. This capacity supports training and inference, but hardware power alone does not establish the fidelity of a particular translation.

Reference documentation: Slator: Inside DeepL · DeepL: About its language model · DeepL: How does it work? · DeepL and NVIDIA infrastructure.

Languages, variants and Catalan

The translator’s current language list includes Catalan, Spanish and English. Not every combination supports the same features: glossaries, formality and alternatives have separate coverage. Regional variants also exist; Latin American Spanish is associated with the next-generation model.

For a multilingual project, a useful check is a matrix of source language, target language, format and required feature. A language appearing in Translator does not mean it is available in Write or every voice offering. Language should not be equated with country either: the recipient may need professional vocabulary or a variant different from the one suggested by their location.

According to DeepL’s team, the 2025 expansion required changes to automatic detection and language selection. For short texts, acronyms or mixed-language material, explicitly setting the source language helps prevent the entire operation from starting with an unsuitable detection.

Reference documentation: DeepL: How we launched over 70 new languages · DeepL: Translator languages.

Text, documents and output formats

Access is available through the web, desktop and mobile applications, browser extensions and integrations. The simplest workflow is to enter text, choose a language and review the translation. With documents, the structure of the final file matters alongside its linguistic content.

DeepL documents support for Word, PowerPoint and PDF, among other formats, with different conditions by plan. Excel and additional formats are not enabled in every offering; files translated using a free account may not be editable. The API has its own limits, separate from web use.

For professional delivery, keeping the editable source file usually makes tables, headings and breaks easier to correct. A translation may take up more space than the original: checking layout is a different task from checking meaning. Before processing a batch, testing a document with representative tables, notes and references can reveal problems before they multiply.

Reference documentation: DeepL: File formats · DeepL: About Translator.

Glossaries, translation memory and editorial consistency

A glossary establishes preferred translations of terms and expressions, adapted to context and grammar. A translation memory retains corresponding segments; style rules guide writing conventions. Customization Hub brings these resources together, addressing different levels of language.

A useful terminology base identifies the concept, agreed translation, context, owner and date. For example, an equipment name may remain unchanged as a brand, while its functional description needs translating. Combining both into a single equivalence creates unclear instructions.

Resolve contradictory variants before sharing a glossary. A preference should not become a universal rule when a term changes meaning across activities. Translation memories also need maintenance: reusing a sentence approved years ago does not guarantee that it still describes the current procedure.

Reference documentation: DeepL glossary · DeepL: New tools and DeepL Agent.

DeepL Write: revising the same language

Write suggests improvements to wording, grammar, tone and style. The product page reviewed lists English, German, French, Spanish, Italian and Portuguese; it does not list Catalan. Its purpose is to improve how a text is expressed, whereas Translator changes its language.

A practical sequence is to clarify the original first, translate it and then review the result. Repeatedly rephrasing a sentence can make an obligation sound like a suggestion or remove a condition. For operational communications, accepting changes individually helps preserve the purpose, even when an alternative sounds more elegant.

Reference documentation: DeepL Write product.

Voice, conversations and meetings

Voice began in November 2024 with products for virtual meetings and face-to-face conversations, focused on conveying speech across languages. The April 2026 expansion added speech output, group conversations and an API for integrating spoken translation.

In a meeting, translating while someone speaks means managing incomplete context, turns and delay. A realistic test should include proper names, accents and workplace vocabulary; a demonstration in a quiet room does not necessarily represent a conversation in a production area.

Sabentis proposes using these resources to facilitate questions in multilingual sessions and confirming agreements in writing afterwards. When a message determines an immediate action, agree on a way to confirm understanding and an alternative channel. Voice availability does not establish that the system is suitable as the sole means of emergency communication.

Reference documentation: DeepL launches Voice · DeepL: Voice-to-voice translation.

APIs, assistants and connected workflows

The API receives text and a target language, with options for source language, glossary and context. The context field can guide translation without itself being translated; items in a batch are processed independently. Retaining the relationship between each source identifier and its result is useful.

The MCP connector announced in September 2026 enables DeepL use from compatible assistants, including ChatGPT, Claude and Microsoft Copilot, by linking an account. It can translate text and documents with terminology settings. This adds a provider to the workflow: understand which content each connected service receives.

Translation Flow connects projects with storage and content-management systems and provides tracking and review. DeepL Agent, announced in 2025, handles tasks in business applications. These are different propositions from manually translating a file; before automating, define who reviews, which event starts the work and which action publishes the result.

A maintainable design records the source version, languages, review status and destination. If a process fails halfway through a batch, these details allow it to resume without confusing a pending translation with an approved one or replacing reviewed material with unchecked new output.

Reference documentation: DeepL API: Translate text · DeepL MCP in Microsoft Copilot, ChatGPT and Claude · DeepL Translation Flow · DeepL: New tools and DeepL Agent.

Reception, studies and recognition

Early reception highlighted the naturalness of translations. In 2021, Slator qualified professional assessments by language pair and content. A favourable opinion about fluency does not establish superiority for every task.

Heiss and Soffritti studied Italian-to-German translations using results available in January 2018. They identified educational possibilities and errors requiring post-editing. Their work is an earlier contribution to evaluation and training, not a measurement of the 2026 service.

WMT24++, published in 2025, compared systems including DeepL across 55 languages and variants using automatic metrics. LLMs led its results; the authors left human evaluation for subsequent work. Traditional translations were collected in September 2024.

DeepL received 2020 Webby awards for Best Practices and Technical Achievement. Its September 2025 announcement records inclusion in Forbes AI 50 and Cloud 100. Recognition describes reception and business standing; it does not certify a translated instruction.

To support a reasoned choice, Sabentis proposes evaluating your own sample and classifying errors by effect: a stylistic preference, an incorrect term and a missing negation do not have equal importance. Time savings should be measured after review, not only until the first draft is produced.

Reference documentation: Slator: Inside DeepL · Heiss and Soffritti: DeepL in translator teaching, 2018 · Deutsch et al.: WMT24++, 2025 · The Verge: Webby Awards 2020 winners · DeepL: Forbes recognition 2025.

Data and service conditions

The policy distinguishes the free Translator and Write web services, which may use content to improve models and do not permit personal data, from Translator Pro, API Pro and Write Pro, whose texts are not used for that improvement. Processing personal data in Pro requires justification and appropriate contractual conditions.

Within an organisation, distinguish publishable documents, internal information and identifiable information. Removing visible names does not guarantee anonymity when context identifies a person. A corporate account authorised to translate text also does not automatically authorise connecting other systems or transferring their data.

Access reviews should cover deliberately stored resources, such as glossaries and translation memories, alongside the text of a one-off translation. In conversations, participants and administrators need to understand the intended use of audio and transcripts. Decisions should concern the contracted product, rather than a general privacy claim about the brand.

Reference documentation: DeepL: Privacy policy.

Applications in occupational safety and health

It can help prepare understandable prevention information for multilingual teams. In occupational safety, translation must preserve duties, prohibitions, conditions and units. A fluent version needs checking by someone who understands both the language and the work activity, including the relevant technical terminology. Readability and technical accuracy are separate checks.

One application is translating induction materials while maintaining terminology for jobs, equipment and locations. Another is preparing survey or communication versions for subsequent linguistic review, without assuming psychometric equivalence of questionnaires. It can also support an initial reading of foreign-language technical documentation. These scenarios proposed by Sabentis require distinguishing translation for understanding from translation intended to guide actions at work.

Worked example and checks

For example, a team translates induction instructions for workers from several countries. A reviewer compares every step with the original, maintains a glossary of critical terms and checks with an intended reader that the explanation makes sense in the actual work context.

1. Start with a current, editable instruction with an identified version and owner; remove unnecessary internal comments.
2. Create a reviewed glossary for equipment, locations, proper names and concepts whose translations are not interchangeable.
3. Translate a representative section before processing the whole document.
4. Compare the original and output, checking negatives, conditions, numbers, units, sequence and references to illustrations.
5. Inspect the file visually and check understanding with intended readers before approving its use.

If a sentence allows two interpretations, correct the original first and translate it again. Recording terminology decisions in a bilingual table prevents reviewers from introducing different variants. Acceptance means more than natural-sounding prose: the translation must retain who performs each action, when and under which conditions, without adding or removing instructions. Keep the reviewed translation linked to the source version so later source changes trigger a deliberate language review.

To keep versions aligned, assign each translated section the identifier of its original. When a condition or figure changes, review that section in every language and check cross-references. This discipline prevents a linguistically correct translation from circulating after its source instruction has been replaced or expanded. Record who checked terminology and which source version was used, without implying that machine translation itself provides that approval.

Quality evaluation and boundaries

DeepL describes a Translation Quality Evaluation project using a model separate from the translator: it compares segments and flags possible errors with severity levels for human review. Its documentation describes development, an API and access to a test environment; this evaluation should not be assumed to accompany every ordinary translation.

The main practical limitation is that incorrect text can sound entirely natural. Checking spelling and style alone does not necessarily detect omissions, unclear references or changes in scope. Back-translating into the original language may provide clues, but two systems can repeat the same error and create a false sense of agreement.

Checks should match the use: general information for reading, text intended for publication and a document directing a task require different criteria. For critical content, recording discrepancies and their resolution provides more traceability than a general approval of the entire file.

Reference documentation: DeepL: Translation Quality Evaluation research.

Limitations and conditions of use

Machine translation may alter negatives, abbreviations or ambiguous technical terms. Tables and formatting also need review. Unreviewed machine translation should not be presented as a certified translation, legal validation or guarantee of understanding.

An incorrect glossary can propagate an error consistently. Preserving file format does not guarantee readable tables, page breaks or captions. Confirm the specific language and function before committing to a delivery, because language support differs between products. Before uploading internal information, review the contracted service and its data-handling conditions.

Reference documentation: DeepL: File formats.

Choosing and complementing the tool

Choose DeepL when the main task is translating existing content with controlled terminology. Microsoft Word complements comparison, editing and approval of the final document. Grammarly can support language review where compatible, keeping style correction separate from technical verification. No combination removes the need to retain a current original and a record of revisions.

Frequently asked questions

Does a glossary replace technical translation review?

No. It helps maintain consistent terms but does not validate relationships between sentences, conditions or complete instructions.

Can I translate any file?

Formats and limits are specific to the product and plan. Changing an extension does not turn an incompatible file into a valid one.

Does translation automatically preserve a questionnaire’s validity?

Do not assume so. Linguistic equivalence and suitability of the instrument for a new context are separate questions.