New Open-source Inference Model Qwen3: Mathematically Strong, Code Slick, and Context Expandable to 100,000

TechWeb news on July 31st, the Qwen3 team published a message on their public platform, announcing that Qwen3-30B has achieved a major upgrade, with the new inference model Qwen3-30B-A3B-Thinking-2507 officially released. The new model has expanded thinking abilities, improved reasoning quality and depth, and is a more intelligent, agile, and all-round new inference model.

The new model has made significant improvements in reasoning ability, universal ability, and context length. In the AIME25 evaluation focusing on mathematical abilities, it scored 85.0 points, and in the code ability test LiveCodeBench v6, it achieved a score of 66.0, surpassing Gemini2.5-Flash (thinking) and Qwen3-235B-A22B (thinking). The new model's knowledge level (GPQA, MMLU-Pro) has also made significant progress compared to the previous version; in writing (WritingBench), agent ability (BFCL-v3), multi-round dialogue, and multilingual instruction following (MultiIF) evaluations, Qwen3-30B-A3B-Thinking-2507 has surpassed Gemini2.5-Flash (thinking) and Qwen3-235B-A22B (thinking). The original support for 256K tokens can be expanded to 1M tokens.

It's worth noting that the new model has increased its thinking length, and the team recommends setting a longer thinking budget when handling complex inference tasks, which can fully unlock the potential of the new model.

Qwen3-30B-A3B-Thinking-2507 is now open-source on MagCAD community and HuggingFace, with Qwen Chat synchronization online.