DeepL Beat Google at Translation by Refusing to Be a General Model

Jarek Kutylowski, cofounder and CEO of DeepL, joined Redpoint partner Jacob Effron on Unsupervised Learning after the German company raised at a $2 billion valuation and crossed 100,000 business customers. DeepL had been training language models for translation long before ChatGPT made that work fashionable. The day the chatbot landed was still a jolt inside the company—not because the science was new, but because the public finally felt how magical the technology could be. Kutylowski says that moment did more to educate people about AI, and about what DeepL had been building, than years of academic papers.

The company has been up against Google Translate since day one, and Kutylowski has written that contest into the culture. Competition, he argues, produces the best outcome for users, and keeping giant rivals in mind creates a permanent sense of urgency. He does not think AI translation would sit where it does without DeepL entering the category. Large language models, arriving later, let the product feel less like a document in and a document out, and more like a colleague sitting next to you, perfecting the draft.

The business case starts with a pair of numbers. Only 18 percent of the world speaks English, and even in the United States, the largest English-speaking market, 20 percent of people speak another language at home. DeepL's mission is to help companies communicate across those languages so operations run better internally and new markets become reachable. The industries range from law and manufacturing to retail, media, and government, but the work that shows up most is high-value: cases where accuracy and quality are worth paying for.

On the external side, Nikkei—the Japanese media group that owns the Financial Times—uses DeepL to translate thousands of articles among Japanese, English, and Chinese, reaching readers who would never have seen the original. Internally, the pattern is often a multinational in a regulated industry: a German contract lands on the desk of a specialist sitting in the United States. A Japanese carmaker Kutylowski works with keeps all of its R&D in Japan while customers sit in Europe and the US; customer feedback has to flow back into the product roadmap, which only works if those offices can talk without friction.

Effron puts the question directly. Look back at the original Transformer paper, whose running example was French-to-English translation, and explain how Google did not win the category. In 2017, Kutylowski says, plenty of investors assumed the incumbent would simply figure it out. His short answer is a big fuss to the market—get there first—and then keep innovating at the same pace. DeepL launched with quality that already beat the general-purpose tools of the day, and it has not let the cadence slip.

The longer answer has three layers. DeepL built academic-level research inside the company and specialized on one use case: high-value translation for businesses, a job that is Google's among many and DeepL's entire company. Just as important, the team was founded in Europe. Kutylowski lives a half-hour drive from Belgium, France, and the Netherlands, so the people writing the models are surrounded by languages they actually have to use.

That geography, he says, makes the team understand the problem and stay motivated to solve it. Engineers and researchers should not underestimate the power of working on something that is meaningful in their own lives. The same competitive DNA that started with Google Translate still runs the company: huge players sit in the market, and remembering them is a driving force. The user, not the incumbent, is who has to win.

Asked where DeepL does original research versus riding the giant general models, Kutylowski does not hesitate: it is a build-it-yourself company. In the early days there was no other option—models, tooling, even data centers simply were not available off the shelf. He still believes ownership of the whole vertical stack, from go-to-market through product and engineering down to training and architecture, lets a team trickle a customer problem all the way into the weights. Prompt engineering, he says, cannot reach those levers.

When a new general model lands, DeepL does not reach for a single go-to prompt. It keeps test sets and asks what the user is trying to achieve—maximum accuracy or maximum fluency. Those trade-offs are baked into DeepL's own models by default. They try to figure out whether the text is technical and should be fully accurate, or marketing copy that should sound native.

The product example he reaches for is terminology. Customers need to lock the language of product manuals and internal jargon, and DeepL has been running glossary control for years in a way no other translation provider has matched. It is not a find-and-replace. Swap a word and the grammar around it has to change; a customer may insist that A always become B, except that A sometimes has two meanings and a blind rule would be wrong. That feedback loop from production into the model is what a prompt-only shop never gets.

Under the hood there is not one giant model. DeepL runs different models by language pair, because German-English has far more parallel data than Polish-English, so the parameter counts differ. Similar languages get grouped so the models can learn from each other, and some models are kept smaller purely to save inference compute at the scale DeepL already operates. Effron sums it up: you do not need a model that can write Shakespearean sonnets to translate a scientific paper. Kutylowski agrees.

Outsiders assume an AI translation company no longer needs human translators. DeepL's practice is the opposite. Human data has been growing in influence, Kutylowski says, and will only get more important as reinforcement learning for language models matures. The company has been running large annotation projects internally for years—the same kind of work Scale AI turned into a business—using professional translators both to show the model what a good translation looks like and to keep quality assurance tight.

Specialized models face a bar general chatbots do not: customers expect high quality, and they expect it to be steady. That means the QA has to be top-notch, and it means hiring people who speak the language extremely well. DeepL recruits across the globe so that Brazilian Portuguese is labeled by a Brazilian Portuguese native speaker. The hiring is exciting and also complicated, and Kutylowski treats it as a central, lasting part of the company rather than a temporary bootstrap.

He is not dogmatic about keeping every labeling task in-house. DeepL is already thinking about outsourcing parts of it. The real variables are how much control you need and how sensitive the task is to quality. Some jobs want a huge pile of data at a merely decent bar; others want a small set of the best possible quality, down to choosing the specific individuals who will do the work.

That last requirement is why proximity matters. DeepL has watched annotators who performed extremely well fall sick for a week, watched quality drop for that week, and watched it recover when they returned. If you stay close to the process, you can catch those swings and still have the data your tests require. High-quality specialized data is now the scarce input, and finding those people is a large hiring endeavor of its own.

Effron voices the anxiety a lot of builders feel: you can spend years plugging holes that the next general model will fill, so maybe you should sit on a beach until GPT-5 arrives. Kutylowski's reply is blunt. For a company that wants to be cutting-edge, that train of thought is dangerous. Innovation means throwing away a lot of results—some deprecated by a competitor's model, some by your own research when a better idea shows up. He describes it as founder mode on the research side: you have to hold the conviction.

Failed attempts are not waste. Every dead-end architecture, he says, buys a better understanding of the problem, and you almost never jump from here to there without failing a few times in between. The tedious steps are the path. There is a temptation to be financially responsible, cut those investments, and drink coffee in Hawaii. He does not believe you can build an innovative company that way.

The large failed experiment he is willing to name is custom models for individual customers. Before AI translation, and in its very early years, that was the industry's only way to hit certain quality levels. DeepL tried to productize it—scalable, repeatable, without an AI engineer on every account. What they learned is that out-of-the-box translation quality has surpassed the gains custom models could add, and the extra hassle does not scale.

OpenAI's o1 reasoning model, freshly released at the time of the taping, landed as a slightly less exciting update from DeepL's chair. Reasoning helps because you have to understand the world in order to translate well, but reasoning is not the core of translation itself. The capability Kutylowski is actually waiting for is not a flashier general model. It is the ability to do more with less compute.

Asked what model capability would be a game-changer, Kutylowski does not name a flashy feature. He wants to do more with less compute. Pouring ever more compute into models is producing results, but it feels like brute force, and he would rather make the models a little smarter. Transformer was the proof that architecture can move capabilities as much as compute can. The giant general-model labs, he argues, are not necessarily incentivized to invest in smaller models, because they hold a monopoly on huge compute.

Figuring out how to do slightly more with slightly less falls to specialists like DeepL. His other infrastructure wish is simple and widely shared: he wishes more people were building GPUs, or GPU-like solutions. At the scale DeepL already runs, that scarcity is not abstract. Advice for teams heading toward the same loads starts with a placement choice. Be conscious about where you run the work.

Hyperscalers are a great kick-start, especially early and in the current market. Once you hit scale, think about your own data centers. DeepL has operated them from the beginning because nothing else was available then; the options are different now, but cost efficiency still favors owning the iron for massive research and inference loads, including the free product. Own facilities also get you onto the newest hardware faster, which matters when you want to train on whatever just shipped.

GPU tooling, he says, is still early, and the compute is rare. CPU clouds could afford layer upon layer of abstraction because CPUs were cheap; GPUs are scarce and powerful, so you have to optimize for both environmental and commercial sustainability. The hidden cost of owning the iron is development speed: running your own is much more complicated. That is why DeepL is moving large parts of the stack toward a hybrid cloud and keeping on-prem only what cost, efficiency, or customer data-protection rules require.

People who do not live in translation sometimes treat it as a solved problem. Kutylowski's first correction is that quality still depends on the language pair. The most-used pairs work well because they have the most training data and attract the most research attention; lower-resource pairs are a different story. More data still means better translation, the same way it does for models in general.

Use cases split into tiers. One-to-one communication—sending an email—is, for the best-resourced languages, solved very well, and a newspaper can already publish machine-translated articles without human supervision after the fact. A billion-dollar company's marketing website still wants someone in the loop. An operating manual for a nuclear power plant wants more than a quality checker: it wants a person who will take responsibility and accountability for the final result. That last point is, for him, the generic problem AI creates. The technology unloads a lot of work, but someone still has to be responsible at the end.

Evaluation has been a long journey of its own. When DeepL started in 2017, synthetic metrics like BLEU were state of the art, and the company quickly reached a quality level those scores could not capture—a problem the rest of AI later inherited. Automatic metrics still help as a rough orientation during training, and they have gotten better, but the real test is always human. Thousands of translators judge next-gen models and competitors on accuracy, nuance, and how native they feel, always comparatively. An absolute score of 97 versus 85 is too unspecific to mean anything.

On moats, he splits the world in two. The long tail of use cases is not worth a Blackwell cluster; those teams should build on standard technology. Translation is indisputably one of the super-big use cases, where specialized models and a vertically integrated, ready-made product make sense. Customers tell DeepL they do not want to run POCs internally: they want something they plug in, that employees already know from the consumer product, and just roll out. Better products have always won in software; general AI that can do everything is, in practice, pretty complicated to get done reliably.

Quality on smaller language pairs still has room to run, but the next frontier Kutylowski names is spoken language and voice. Good AI text translation has already changed the language-service industry and how people consume foreign content. It has not yet changed conversation. He and Effron are still doing the podcast in English rather than Kutylowski speaking Polish. Demos from the big labs exist, and DeepL is running its own research; turning that into a product you can integrate everywhere is what excites him. If real-time synchronous speaking models arrive, he thinks business will start to ignore location: a founder in Cologne could take cofounders in Tokyo and the US and each could speak their own language.

At company scale, employees in every language would get access to education, learning resources, and knowledge without the old barrier. Personal connection and culture will still make learning languages valuable, and he is still asking his own children to learn them. Better translation might even make local languages more important, because people can reach the global ones without everyone having to speak English. You still would not want to talk to your partner through a phone for twenty years of a relationship. Put to a 30-year stake, he guesses the average person will speak fewer languages, but with more fun, the way people still learn chess after AI became world champion.

The timeline is a few years for something usable, longer to perfect, because speech is harder than text. It is stream-based rather than chunked into sentences, people speak carelessly, and even a clean conversation transcript is unstructured in a way written language is not. Latency matters, but the deeper problem is that models will have to be taught to translate a different kind of language. Adjacent to that, he is excited about language-learning apps: speaking well used to require an in-person teacher and was a privileged thing to do. A model that hallucinates a bit is fine in a practice dialogue; he only wonders whether talking to a phone will ever be as enjoyable as learning over dinner.

The quickfire round compresses the argument. Overhyped: general models. Underhyped: specialized models, where the value is being created. The biggest surprise in building DeepL was beating the tech giants in 2017; the engineer's mistake was assuming technology alone would be enough, without product and commercial motion. He has become a Radical Candor convert, feels kinship with Perplexity taking on Google Search, and would go into medicine if he had not started DeepL—drug discovery and access to doctors, he thinks, would change lives even more than translation. Hardware, for an engineer, is cooler than software. The company grew by word of mouth among working translators, who let AI take the mundane part and add value where it still needs a human. Anyone can try DeepL Translate and DeepL Write for free at deepl.com, with no signup.