Qwen3-2507

What it is

Qwen3-2507 is a refreshed release of the Qwen3 family of open-weight language models, published in stages between 21 July and 6 August 2025. Its first member, Qwen3-235B-A22B-Instruct-2507, was released on 21 July 2025 as an updated non-thinking version of Qwen3-235B-A22B.^[1] The models are sparse mixture-of-experts networks: the 235B model card lists 235B total parameters, of which 22B are activated, across 94 layers and 128 experts with 8 experts activated at a time.^[2] Attention uses 64 query heads and 4 key-value heads.^[2] The weights are released under the Apache 2.0 licence.^[2]

Separate models instead of hybrid thinking

The earlier Qwen3 checkpoints were hybrid: one model could answer directly or produce a reasoning trace before answering. The 2507 line dropped that arrangement in favour of separate models. The Instruct model supports only non-thinking mode and generates no `<think>` blocks, while reasoning is handled by a Thinking counterpart released as its own model.^[2] The practical effect is that the choice between fast direct answers and long reasoning is made when a model is picked for a job, rather than being switched inside a single checkpoint.

Release timeline

The line was rolled out model by model. Qwen3-235B-A22B-Instruct-2507 came first on 21 July 2025.^[1] Qwen3-235B-A22B-Thinking-2507 followed on 25 July, the smaller Qwen3-30B-A3B-Instruct-2507 on 30 July, and Qwen3-30B-A3B-Thinking-2507 on 31 July.^[1] The series closed on 6 August 2025 with the Qwen3-4B Instruct and Thinking models, which were described as the final open release of the 2507 line.^[1]

Longer context

Context length was one of the headline changes. The model card gives a native context of 262,144 tokens, extendable to about 1 million, listing a maximum of 1,010,000 tokens.^[2] On 8 August 2025 the Qwen team said the 235B and 30B 2507 models could handle 1M-token inputs.^[1] A context that large changes what can be put into a single request: whole code repositories, long reports or multi-document sets can be supplied together instead of being retrieved and trimmed in advance, though very long inputs still cost proportionally more computation.

Benchmark results

Against the original non-thinking mode of Qwen3-235B-A22B, the Instruct-2507 model posts much higher scores. On MMLU-Pro it reaches 83.0 versus 75.2, on GPQA 77.5 versus 62.9, and on AIME25 70.3 versus 24.7.^[2] On LiveCodeBench v6 the score is 51.8 versus 32.9, on Arena-Hard v2 79.2 versus 52.0, and on SimpleQA 54.3 versus 12.2.^[2] The improvements cited in the model card cover instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool use.^[2] The gaps are uneven: they are moderate on knowledge-heavy multiple-choice tests such as MMLU-Pro, and very large on competition mathematics, coding and short factual question answering.

Where it sits in the Qwen3 series

Qwen3-2507 is best understood as a mid-series revision rather than a new architecture. The parameter layout, the expert count and the attention head configuration stay the same as the 235B model they are built on, and the licensing is unchanged.^[2] What changed is the training recipe and the packaging: sharper results in reasoning-adjacent tests, a much longer usable context, and a clean split into Instruct and Thinking variants across several sizes, from the 235B flagship down to the 4B models that ended the run.^[1] The activated-parameter design keeps the compute per token at the level of a far smaller dense model, which is what makes a 235B-parameter release practical to serve at all. For users, the 2507 line marks the point where the Qwen3 family stopped being one hybrid model and became a menu of specialised ones.