DeepL SE, a German developer of artificial intelligence translation software, is reportedly considering to go public on a U.S. stock exchange.
The company received a $2 billion valuation following a $300 million funding round last March. Citing sources familiar with the matter, Bloomberg reported today that DeepL is seeking a market capitalization of up to $5 billion with its initial public offering. The steep valuation increase hints the company may be experiencing the same kind of rapid revenue growth as larger rivals such as OpenAI.
DeepL’s flagship offering is DeepL Translate Pro, an AI model that can translate documents into more than 30 languages. The company claims that the algorithm provides better accuracy than competing algorithms. In an internal comparison with Google Translate, DeepL determined that users were 30% more likely to prefer its model. That percentage rose to to 130% in an assessment that involved Microsoft Corp.’s translation software.
DeepL’s website states that its translation model was developed with “human-guided training methodology.” That hints the company may be using a training method such as RLHF, or reinforcement learning from human feedback. It’s one of the most popular ways to increase the output quality of language models.
In a RLHF project, developers ask users to share their preferences for how an AI model should respond to their prompts. Those preferences are then encoded into a second AI model. That second algorithm trains the first AI to take into account user preferences when generating output.
DeepL sells DeepL Translate Pro alongside two other models. The first can translate speech in near real-time during Zoom and Microsoft Teams meetings. DeepL says that the algorithm supports more than 30 languages.
AI models sometime translate technical terminology and company-specific jargon in unexpected ways. To address that issue, DeepL enables organizations to provide its software with custom glossaries. Those glossaries can specify how DeepL models should translate business terms.
The company says its algorithms lend themselves to a wide range of use cases. A retailer with an international presence could use it to track new business regulations in each market where it operates. A marketing teams, meanwhile, can have DeepL localize their ads and other promotional content.
DeepL’s third offering is a text generation model. DeepL Write, as it’s called, can point out grammar, punctuation and spelling mistakes. It also performs more advanced tasks such as rewriting paragraphs.
According to Bloomberg, DeepL has held preliminary discussions with potential financial advisers about its planned listing. The IPO is expected to take place as early as next year.