Qwen3-2507 was released in July 2025, splitting thinking and non-thinking into two independent models, extending the context to 256K, and separately releasing the programming-specialized Qwen3-Coder and the translation-specialized QwenMT.