Qwen3-30B-A3B-Thinking-2507

Qwen3-30B-A3B-Thinking-2507 is the July 2025 refresh of Qwen's 30-billion-parameter mixture-of-experts reasoning model, part of the Qwen3-2507 update that split the original hybrid thinking models into dedicated Instruct and Thinking variants. The 2507 refresh shipped in three sizes, 235B-A22B, 30B-A3B, and 4B, with the thinking line focused on deeper reasoning for mathematics, science, coding, and other expert-level tasks. Qwen describes it as achieving state-of-the-art results among open-weight thinking models.

The 30B-A3B configuration activates roughly 3 billion parameters per token, making it far cheaper to serve than the 235B flagship while keeping the full 256K native context window, extendable to 1 million tokens. Like all Qwen3 models it ships under Apache 2.0, and it is available on Hugging Face as Qwen/Qwen3-30B-A3B-Thinking-2507. Exact per-size release dates within the July 2025 refresh are era-level; the refresh was announced in the third week of July 2025.