Qwen updates Qwen3 with separate instruct and thinking models

Background

The original Qwen3 release, in April 2025, combined thinking and non-thinking behaviour in one model. The 2507 update changed that design.^[1]

What happened

On 21 July 2025 the Qwen team released Qwen3-235B-A22B-Instruct-2507, an updated non-thinking version of Qwen3-235B-A22B. Thinking versions followed in the next weeks: Qwen3-235B-A22B-Thinking-2507 on 25 July, Qwen3-30B-A3B-Instruct-2507 on 30 July, Qwen3-30B-A3B-Thinking-2507 on 31 July, and the Qwen3-4B Instruct and Thinking models on 6 August 2025.^[1]

Key details

The Instruct model supports only non-thinking mode and generates no think blocks. This is the main structural change from the original Qwen3.^[2]

The team's comparison with the original model's non-thinking mode showed higher scores on several tests. MMLU-Pro rose from 75.2 to 83.0, GPQA from 62.9 to 77.5, and AIME25 from 24.7 to 70.3. LiveCodeBench v6 rose from 32.9 to 51.8, Arena-Hard v2 from 52.0 to 79.2, and SimpleQA from 12.2 to 54.3.^[2]

The improvements were cited in instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool use. The release was under Apache 2.0.^[2]

Several terms in the update need explaining. An instruct model is tuned to answer directly. A thinking model is tuned to produce a reasoning process before its answer. The 2507 line is built from these two kinds of model, so a user can choose the one that fits the task. The think blocks mentioned above are the sections in which a thinking model writes out its reasoning.^[2]

The benchmarks in the comparison test different abilities. MMLU-Pro is a harder version of a broad knowledge test, and GPQA tests difficult science questions. AIME25 is a set of competition mathematics problems. SimpleQA tests factual questions, and LiveCodeBench v6 is a programming test.^[2]

Impact

The update made instruct and thinking versions separate products within the same family, so a user could pick the version that fit the task.