BABA(US:BABA) 量子位·2025-07-23 00:24Core Viewpoint - The article highlights the launch of Qwen3-Coder by Alibaba's Tongyi team, which has set a new state-of-the-art (SOTA) in AI programming, surpassing both open-source and closed-source models in the industry [1][3]. Group 1: Product Features - Qwen3-Coder includes multiple versions, with the strongest being Qwen3-Coder-480B-A35B-Instruct, featuring a 450 billion MoE model and 35 billion active parameters [5]. - It natively supports a context length of 256K and can be extended to 1 million using YaRN technology [6][23]. - The command-line version, Qwen Code, has been developed based on Gemini Code, allowing for prompt and tool invocation protocol adaptation [8]. Group 2: Performance and Capabilities - Users have reported impressive results, such as creating interactive animations and dynamic weather cards with simple prompts [11][13]. - The model can easily generate a playable Minesweeper game and an editable resume template, showcasing its versatility [16][19]. Group 3: Technical Details - During the pre-training phase, Qwen3-Coder utilized various scaling techniques to enhance model capabilities, with training data totaling 7.5 trillion tokens, 70% of which is code data [22]. - The post-training phase involved scaling code reinforcement learning (RL) to improve the model's performance on real-world coding tasks, achieving high success rates in code execution [24][27]. Group 4: Open Source vs Closed Source - Qwen3-Coder is open-source under the Apache License Version 2.0, making it commercially friendly and empowering developers [29][30]. - The article emphasizes that this release represents a significant leap for open-source programming agents, positioning Chinese models at the forefront of the industry [34].