At Rise of AI 2025 in Berlin, Stefan Mesken, Head of AI at DeepL, explains how DeepL has built an AI-native, vertically integrated company since its 2017 founding in Cologne. He describes DeepL's mission as breaking down language barriers, and argues the company's advantage comes from owning its infrastructure, data access and product development rather than being a generic model provider. He walks through DeepL's product history, from the original translator through the Glossary, DeepL Write, DeepL Voice and the Clarify feature, and explains the 2022 decision to move from custom model architectures to LLMs, which led to the DeepL Mercury supercluster built with Nvidia. He outlines guiding principles: staying focused on the mission, delivering solutions rather than raw technology, avoiding benchmark chasing, and treating failures as information about the next S-curve of progress. In a short Q&A he discusses linguistic surprises in Chinese and Japanese translation and the challenge of translating culturally loaded political terms such as digital sovereignty.
Key takeaways
- DeepL is vertically integrated, owning its own infrastructure, data access and product development rather than acting as a generic model provider.
- DeepL does not publish scientific papers, keeping research tied to product development and customer impact rather than benchmark scores.
- The 2022 decision to move from custom model architectures to LLMs led directly to building the DeepL Mercury supercluster with Nvidia.
- DeepL released next-generation models alongside older ones so customers could weigh quality gains against slower inference before switching.
- Clarify has the AI flag passages where it lacks context and ask the user to supply missing information instead of guessing.
Detailed notes
DeepL as a vertically integrated AI company
Mesken introduces DeepL as an AI-native company founded in 2017 in Cologne, Germany, competing with hyperscalers in language AI. He says DeepL is vertically integrated: it owns its own infrastructure, its own data access, and builds its own products, which he calls the reason he joined the company rather than a research lab remote from real-world impact. He states DeepL's mission is simple, breaking down language barriers, and says every decision is filtered through that mission, including deliberate decisions not to pursue image generation or become a generic model provider for others to build on.
From translator to Glossary, Mercury and Clarify
He walks through DeepL's product timeline. In 2017 DeepL launched its text translator, aiming for quality where translated text is indistinguishable from native text. The Glossary followed as an attempt to capture business-specific context. In 2023, DeepL Mercury, described as the first H100 supercluster in Europe, went online in a data center in Sweden, giving DeepL frontier-model training capability. In 2024 DeepL launched DeepL Write, and later DeepL Voice became generally available, followed by the Clarify feature, which he calls first of its kind in the language AI ecosystem.
Guiding principles: solutions over technology, benchmarks over hype
Mesken lays out principles behind DeepL's development approach: knowing the mission and saying no to unrelated opportunities; building solutions for a customer problem rather than technology for its own sake; and avoiding chasing leaderboards and synthetic benchmarks, which he says rarely capture the real underlying translation problems of fluency and adequacy. He notes DeepL does not publish scientific papers, both to keep research tied to product development and to keep the team focused on customer impact rather than benchmark scores. He describes progress as a series of overlapping S-curves, where DeepL tries to exploit the current curve while watching for the next one.
The 2022 pivot to LLMs and building DeepL Mercury
He describes a 2022 internal diagram that led DeepL to identify two directions for improvement: better human-AI collaboration and content generation beyond pure translation. As a result, in March 2022 DeepL decided to move from its custom model architecture to LLMs, despite uncertainty and missing infrastructure and skills. The company contacted Nvidia to build what became DeepL Mercury. By early 2023 DeepL had ordered its supercluster and shifted 80% of its research initiatives toward LLM development. April 2024 brought DeepL Pro, followed in June 2024 by next-generation models released alongside the previous generation so customers could compare quality gains against latency trade-offs, which informed further work on faster inference.
Clarify: letting the AI ask what it doesn't know
Mesken explains Clarify as an attempt to close the gap between AI translation and how professional human translators work, by having the system flag places in a text where it lacked the information to translate with confidence and asking the user to supply it, rather than silently guessing. He frames this as building a better interconnect between language specialists and the AI system that assists them, calling it an example of extending DeepL's reach into customer context rather than treating translation as pure input-output text conversion.
Q&A highlights
Asked what he learned about language that he did not expect, Mesken describes an early project building Japanese and Chinese translation networks despite not speaking either language, and recalls learning that Chinese uses different measure words depending on the type of item being counted. Asked about translating culturally and politically loaded terms such as "digital sovereignty," which he notes Germans and French use with different underlying meanings, he says DeepL's models now capture intent well enough to render metaphors and concepts differently across languages rather than forcing literal structural matches, and that the remaining challenge is exposing this capability to users without overwhelming them with choice.
Quotes
For DeepL, the mission is very simple, breaking down language barriers.
Stefan Mesken
We try to deliver solutions, not technology.
Stefan Mesken
The ultimate limit is not to provide the best AI translation system, but to provide the best translator period.
Stefan Mesken
Questions answered in this talk
Why does DeepL not publish scientific research papers?
According to Stefan Mesken at Rise of AI 2025, DeepL avoids publishing papers so its researchers stay focused on product development rather than academic agendas, and so the team cares about customer impact rather than benchmark numbers.
What is DeepL Clarify?
According to Stefan Mesken at Rise of AI 2025, Clarify is a DeepL feature that scans a text being translated, identifies places where the system lacked enough information to translate with confidence, and asks the user to provide the missing context instead of guessing.
Why did DeepL build its own AI supercomputer instead of relying on external providers?
According to Stefan Mesken at Rise of AI 2025, DeepL decided in 2022 to move to LLMs and contacted Nvidia to build DeepL Mercury, its first H100 supercluster, because vertical integration of infrastructure, data and product was seen as a strategic advantage over competitors.