DeepL claims its LLM outperforms majors in language translation

Language AI company DeepL has announced the launch of its next-generation language model, powered by a highly specialised LLM technology built specifically for translation and editing. The company said the update represents a breakthrough advancement in DeepL’s Language AI platform for businesses and sets a new industry standard for translation quality and performance.

The new solution combines three capabilities that it says are unmatched in this field:

• A language-specialised LLM: The DeepL solution leverages an LLM uniquely tuned for language, resulting in more human-like translations and writing for a variety of use cases with a reduced risk of hallucinations and misinformation.
• Proprietary data: Unlike general-purpose models that simply train on the public internet, the DeepL model leverages more than seven years of proprietary data specifically tuned for content creation and translation.
• Human model tutoring: The DeepL model leverages thousands of language experts specifically trained to ‘tutor’ the model in translation.
   

According to DeepL, the new solution significantly raises the bar for AI translation quality, with blind tests showing the DeepL translation is preferred by language experts 1.3 times more often than Google Translate, 1.7 times more often than ChatGPT-4 and 2.3 times more often than Microsoft.

“This is only the beginning of DeepL’s superior LLM-powered Language AI solutions for businesses,” said Jarek Kutylowski, CEO and Founder, DeepL. “We are a research-driven company to our core and are investing heavily in pushing the boundaries of translation and writing quality, efficiency and adaptability. Our goal is to empower our customers worldwide with the industry-leading technology they need to thrive and scale globally without a language barrier standing in the way.”

“Bringing our new LLM to market is a significant milestone for our research team; however, the real proof lies in the tangible impact we can bring to our customers,” said Stefan Mesken, VP Research, DeepL. “When testing the new model against competitors and our classic model, professional translators noted that Google requires twice the edits and Chat GPT requires three times more edits to get to the same quality. For businesses who translate a large volume of documents, this truly impacts their efficiency and productivity, meaning team members can focus on tasks with high business value.”

The introduction of DeepL’s new, next-gen model represents a significant enhancement to DeepL’s translation services. Blind tests conducted with leading linguists found a 1.7-times improvement with the new LLM, against the old model, for combinations involving English to Japanese and Simplified Chinese. Additionally, there was a 1.4-times improvement for the combination of English and German. These significant accuracy increases will save business users valuable time on manual editing, enabling them to communicate more efficiently across teams and markets.

Translations powered by the next-gen model are now available for DeepL Pro customers for translations in English, Japanese, German and Simplified Chinese, with additional languages coming soon.

Image credit: iStock.com/gesrey