Qwen3-30B-A3B-Instruct-2507

Qwen3-30B-A3B-Instruct-2507 is the non-thinking counterpart in the Qwen3-2507 refresh, the July 2025 update that separated Qwen3's hybrid thinking into dedicated Instruct and Thinking lines. The instruct variant targets instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage, with Qwen reporting substantial gains in long-tail multilingual knowledge and better alignment with user preferences on open-ended tasks compared to the original April 2025 Qwen3 release.

It shares the 30B-A3B mixture-of-experts architecture with about 3B active parameters and the 256K native context window extendable to 1 million tokens. Released under Apache 2.0 on Hugging Face as Qwen/Qwen3-30B-A3B-Instruct-2507, it is the practical default for chat, agents, and RAG workloads where instant answers beat deliberative reasoning. Per-size dates within the July 2025 refresh are era-level.