Since its launch in 2017, DeepL has successfully rivalled machine translation giants such as Google Translate. Today, with the launch of its new large language model (LLM), the startup aims to raise industry standards even higher.

In alignment with the unicorn’s technology development strategy, the new LLM is specifically built for translation and editing. According to DeepL, this fine-tuning to language is what enables it to produce more human-like translations and reduce the risk of hallucinations.

Unlike general purpose models, the LLM doesn’t rely on the internet for training. Rather, it uses DeepL’s proprietary data, tailored for content creation and translation. Language experts also took part in the AI model’s training, tutoring it in terms of translation quality.

DeepL claims its new LLM outperforms models of competitors such as Google, OpenAI, and Microsoft. In blind tests, professional translators said that Google Translate requires 2x more edits and GPT-4 3x more edits to reach the same translation quality as the DeepL model.

Some tests of our own here at TNW (using English-French and English-Greek) had DeepL’s classic model already come out on top against Google Translate.

The company also conducted blind tests between the new and the classic LLM it uses. They demonstrated a 1.7x improvement with the new one for combinations between English, Japanese, and Simplified Chinese. There was a 1.4x improvement in the English-German pair.

Translations powered by the new LLM are now available for DeepL Pro users in four languages: English, German, Japanese, and Simplified Chinese. Users can activate the LLM on the web translator by selecting “next-gen model.”

The launch follows a period of significant growth for DeepL. The startup has expanded its portfolio of languages, opened its first office in the US, and launched new products for businesses, such as the DeepL Write Pro AI assistant.

In May, the company secured a new investment of $300mn (€277mn), reaching a $2bn valuation. It covers 32 languages and counts over 100,000 business users.