Qwen-MT

Qwen-MT is Alibaba's dedicated machine translation model, built on the Qwen3 architecture and fine-tuned with reinforcement learning specifically for translation quality rather than general chat. Launched in July 2025, it supports 92 languages covering over 95 percent of the world's population, with production features like terminology intervention, domain prompting, and translation memory.

The open-weight variants come in 4B, 8B, and 30B-A3B MoE sizes under Apache 2.0, with the 30B-A3B activating only about 3B parameters per token for efficient inference. Alibaba's API tiers (Flash, Plus, Turbo) price output tokens from around half a dollar per million. Benchmarks reported by Alibaba place it ahead of GPT-4.1-mini and Gemini 2.5 Flash on translation tasks and competitive with GPT-4.1, though those are vendor-reported figures. A sibling model, Qwen-MT-Image, extends the line to translating text inside images.