In this article, we will look at what Roblox built and the trade-offs they made.

ByteByteGo provides tutorials, articles, and resources for learning and mastering the Go programming language, covering topics such as syntax, concurrency, and best practices. Developers can learn about Go programming fundamentals, web development with Go, and building scalable applications using Go's powerful features and standard library.

ByteByteGo

Roblox built a single unified transformer-based translation model using Mixture of Experts (MoE) architecture to handle all 256 language direction pairs across 16 languages, replacing the need for 256 separate models. To meet a 100ms latency ceiling at 5,000+ chats per second, they applied knowledge distillation to compress the model from ~1B to under 650M parameters, combined with quantization and model compilation. The serving pipeline includes a translation cache, dynamic batching, and an embedding cache that avoids re-encoding the same source message for multiple target languages. For quality measurement, they built a reference-free quality estimation model that scores translations at word-level granularity without needing human reference translations. Low-resource language pairs were improved via iterative back-translation, and domain-specific slang was handled through human-labeled platform terminology.

How Roblox Uses AI to Translate 16 Languages in 100 Milliseconds

Unblocked: Context that saves you time and tokens (Sponsored)

Making a Billion Parameters Fast Enough for a Conversation