Roblox developed a unified, transformer-based large language model that supports real-time, bidirectional translation across 16 languages with about 100 milliseconds latency. This model is trained on a mix of open-source data, Roblox-specific chat logs, and human-labeled examples, enabling it to understand platform-specific slang and gaming terms. Instead of separate models for each language pair, Roblox uses a single architecture with specialized “experts” for language groups, improving efficiency and translation quality by leveraging linguistic similarities.

The translation system is integrated into Roblox’s in-experience text chat, allowing users worldwide to communicate naturally in their native languages. The platform sends both original and translated messages, letting users view either version. Safety and moderation systems apply equally to translated content. The model is continuously updated with new slang and trending terms through human-in-the-loop feedback and automated evaluation metrics that do not require ground truth translations.

Roblox’s approach has significantly boosted user engagement, session quality, and retention by breaking down language barriers. It also empowers developers and brands to localize content easily, expanding global reach and monetization. Future plans include extending translation to voice chat, images, and 3D content, further enhancing accessibility and platform stickiness. The system is optimized for cost and scalability, supporting over 70 million daily users and billions of messages.