Enterprise Neural Machine Translation at Scale

deepl · Tech · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise neural machine translation

DeepL needed to deliver high-quality enterprise translation across languages and use cases while competing with general-purpose LLMs and large incumbents, maintaining low latency, controlling hallucination/quality risk, and scaling custom context across hundreds of thousands of customers.

Constraints:

Translation must preserve source accuracy while sounding native in the target language.
High-compliance use cases such as contracts require much higher quality than casual messages.
Per-customer fine-tuning is not scalable for hundreds of thousands of customers.
Speech translation requires low latency and specialized handling of noisy input.
GPU infrastructure and scheduling must balance utilization, latency, model size, language pairs, and load.

Architecture

DeepL focuses specialized neural translation models on the translation task, using pre-trained models as starting points, adding substantial compute on top with specialized curated data, building its own GPU/data-center and training/inference frameworks, injecting customer/use-case context instead of retraining per customer, cross-checking outputs against source text, and dynamically scheduling model capacity by language pair and load.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  source-content["Source content"] --> context-injection-layer["Context injection layer"]
  context-injection-layer --> specialized-translation-models["Specialized translation models"]
  curated-training-data["Curated training data"] --> specialized-translation-models["Specialized translation models"]
  specialized-translation-models --> gpu-scheduling-and-inference["GPU scheduling and inference"]
  gpu-scheduling-and-inference --> translated-output["Translated output"]
  human-post-editing["Human post-editing"] --> translated-output["Translated output"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Source content`	Text, documents, contracts, support content, or speech requiring translation.	`Context injection layer`
`Context injection layer`	Supplies customer- or use-case-specific context without per-customer retraining.	`Specialized translation models`
`Specialized translation models`	Neural models tuned for source fidelity and fluent target-language generation.	`GPU scheduling and inference`
`Curated training data`	Specialized data distribution across languages used to train and adapt translation models.	`Specialized translation models`
`GPU scheduling and inference`	Routes requests to available model capacity while balancing utilization and latency.	`Translated output`
`Human post-editing`	Review layer for high-stakes workflows where quality and legal consequences matter.	`Translated output`
`Translated output`	Final localized text or speech translation for enterprise use.	—

Data flow

Source text or speech enters a specialized translation model; customer or use-case context is injected where needed; GPU scheduling routes requests to the right models/language-pair capacity; translation output may be human-reviewed or post-edited in high-compliance workflows.

“how we can inject uh context specific or customer specific information um for for for the particular use case”transcript, stated

“wide range of models like depending on language pairs uh depending on the load that you're getting on the system”transcript, stated

Why this architecture

DeepL argues that translation needs specialized models because the task requires both source-text accuracy and native target-language fluency, and enterprise workflows need quality, context, and latency that generic models or per-customer retraining cannot economically provide.

“you have to both maintain a certain level of accuracy because you the translation needs to be the kind of near to the sort source text”transcript, stated

“this mix of um both uh models that are uh that are good at uh at copying and also at writing itself”transcript, stated

Models used: DeepL neural translation models · Meta Llama pre-trained models · Specialized curated translation models · GPU inference infrastructure

Scale

DeepL launched in 2017.
“We we launched in 2017”transcript, stated
DeepL serves hundreds of thousands of customers.
“with kind of the the hundreds of thousands of of customers that we have”transcript, stated
DeepL started building its own data centers in 2017 because it could not get enough GPU compute.
“in 2017, we we had to start to build our own data centers because we we essentially couldn't get the GPU compute”transcript, stated
Contract and terms translations may be published in 20 different languages.
“translating your terms and conditions and and putting them onto your website in 20 different languages”transcript, stated

Alternatives considered (roads not taken)

Use general-purpose LLMs for translation.

Rejected because: DeepL says specialized models hallucinate less and stay more controlled for translation.

“specialized models and that's kind of one of the differences hallucinate less than general purpose genai models”transcript, stated

Retrain large-scale models per customer.

Rejected because: That does not scale to hundreds of thousands of customers, so DeepL injects customer/use-case context.

“without having to retrain everything. uh basically and with kind of the the hundreds of thousands of of customers”transcript, stated

Use off-the-shelf training/inference infrastructure.

Rejected because: DeepL had to build data centers and frameworks because GPU compute and translation infrastructure were not available enough.

“we had to build our own frameworks for like how do we put the um the training workloads onto the data center.”transcript, stated

Accept high latency for speech translation.

Rejected because: Speech translation quality depends on being as real-time as possible.

“on latency and just making the translation as real time as possible.”transcript, stated

Patterns

`fine-tuning-adaptation` (model-ops)

DeepL starts from pre-trained models but adds specialized translation compute and curated data.

“we're using them them as as pre-trained models. We we're putting still a lot of compute on top of that.”transcript, stated

`context-compaction` (context)

DeepL injects customer/use-case context instead of maintaining separate models for every customer.

“give the models the right context, how we can inject uh context specific or customer specific information”transcript, stated

`latency-optimization` (optimization)

The platform schedules GPU/model capacity by language-pair load while maintaining low latency.

“you want to maintain like low latency for for for your users.”transcript, stated

`human-in-the-loop-gate` (governance)

High-stakes translation workflows still include humans checking and post-editing model output.

“there's still a human in there um checking the translation, post- editing it”transcript, stated

Tradeoffs

Specialize for translation.

Chose Translation-specific architectures and curated data. over General-purpose LLM translation alone..

Rationale: Translation needs source fidelity and fluent native target output.

“we found like there's there's actually better architectures for translation specifically.”transcript, stated

Inject context instead of per-customer retraining.

Chose Customer/use-case context injection. over Large-scale training models per customer..

Rationale: Per-customer retraining is not practical at hundreds-of-thousands customer scale.

“We're trying to find ways on how we can give the models the right context”transcript, stated

Invest in proprietary GPU infrastructure.

Chose Own data centers, frameworks, and GPU scheduling. over Relying only on off-the-shelf compute and inference tooling..

Rationale: Compute access and optimization are central to DeepL's quality and latency edge.

“we also consider that being essential for us to to maintain an edge and and and be able to uh to train those large scale models.”transcript, stated

Gotchas (what bit them)

Translation models can hallucinate when given freedom to write.

Root cause: Models need creativity for fluent target-language output but can make things up when uncertain.

Resolution: Use specialized translation models and cross-check against the original source text.

“sometimes if they don't know what they should be doing they start making stuff up.”transcript, stated

Edge-case inputs confuse models.

Root cause: Half-broken sentences, very short app text, or rarely seen formats are underrepresented in training.

Resolution: Use context and specialized training/evaluation for those input classes.

“halfbroken sentences which due to I don't know some some kind of text parsing uh are slightly weirder”transcript, stated

GPU scheduling for translation is not just generic CPU scaling.

Root cause: Different language pairs and load patterns need different model capacity while preserving latency.

Resolution: Build scheduling technology that reacts to language-pair load by spinning models up or down.

“spin up more models for Japanese because it's the kind of the time zone for for Japanese uh language and then and spin down”transcript, stated

Evaluation

DeepL discusses translation quality by source fidelity, target-language fluency, legal/compliance consequences, post-editing effort, hallucination control, latency, and language-pair/model performance.

Metrics: Translation accuracy to source · Target-language fluency · Post-editing effort · Latency/real-time speech translation · Language-pair quality · Hallucination rate/control

“quality level you are at, you're always unlocking new use cases for being tackled by machine translation.”transcript, stated

“the less edits are necessary, the less changes are necessary. This this really impacts the time uh needed”transcript, stated

Outcomes

DeepL built a translation-focused business around specialized neural models after launching in 2017.
The company serves hundreds of thousands of customers.
DeepL maintains a technical edge through specialized curated data, own infrastructure, and low-latency scheduling.
The source does not provide a single benchmark score, but repeatedly frames quality as the differentiator for enterprise use cases.

The transcript emphasizes enterprise use, hundreds of thousands of customers, language-pair model scheduling, and high-stakes legal/customer-support workflows.

“we're trying to to to provide solutions which just help you uh cross that language barrier”transcript, stated

“depending on on which quality level you are at, you're always unlocking new use cases”transcript, stated