阿里通义千问推出全新推理模型 Qwen3-30B-A3B-Thinking-2507

Core Insights - Alibaba's Tongyi Qianwen has launched a new reasoning model, Qwen3-30B-A3B-Thinking-2507, which shows significant improvements in reasoning ability, general capability, and context length compared to the previous version released on April 29 [1][3]. Performance Metrics - The new model achieved a high score of 85.0 in the AIME25 evaluation focused on mathematical abilities and scored 66.0 in the LiveCodeBench v6 for coding capabilities, surpassing competitors like Gemini2.5-Flash (thinking) and Qwen3-235B-A22B (thinking) [3]. - In various general capability assessments, including writing (WritingBench), agent capabilities (BFCL-v3), multi-turn dialogue, and multilingual instruction following (MultiIF), Qwen3-30B-A3B-Thinking-2507 outperformed Gemini2.5-Flash (thinking) and Qwen3-235B-A22B (thinking) [3]. Context and Deployment - The model supports a longer context understanding, natively supporting 256K tokens and expandable to 1M tokens, which enhances its performance in complex reasoning tasks [5]. - Qwen3-30B-A3B-Thinking-2507 has been open-sourced on platforms like MagicDock Community and HuggingFace, allowing for easy local deployment on consumer-grade hardware [5].