通义千问深夜更新！Qwen3升级版迈向“分离训练”时代，性能全面超越Kimi-K2，Agent能力亮眼

Core Viewpoint - The latest update of Alibaba's Qwen3 model has achieved significant advancements, surpassing top open-source models like Kimi-K2 and even leading closed-source models such as Claude-Opus4-Non-thinking, indicating a competitive edge in the AI large model race [1][3]. Performance Enhancements - The new Qwen3-235B-A22B-Instruct-2507-FP8 model shows remarkable improvements across various core capabilities, including instruction adherence, logical reasoning, text comprehension, mathematics, science, programming, and tool usage, outperforming several leading models in multiple authoritative assessments [3][5]. - In the BFCL (Agent capability) assessment, the Qwen3 model demonstrated exceptional performance, indicating a new level of understanding complex instructions, autonomous planning, and tool utilization [5]. Technical Innovations - The transition to a "separate training" approach marks a significant technological shift, moving away from the previous "mixed thinking mode." This new strategy allows for independent training of the Instruct model for direct responses and the Thinking model for complex reasoning tasks [11][12]. - The Qwen3-235B-A22B-Instruct-2507-FP8 model focuses on "fast thinking," aiming for enhanced speed, accuracy, and strength in tasks like instruction following and knowledge Q&A [12]. Competitive Landscape - The competition in the domestic AI open-source sector has intensified, with each update leading to performance leaps and shifts in leadership among models [14]. - The Qwen3 model has been fully open-sourced on platforms like ModelScope and HuggingFace, allowing AI developers and enthusiasts to experience its capabilities firsthand [15].