Qwen3.8-Maxが16日間でアプリを構築、人間のコーダーを凌駕
alibabaが公開した2.4tパラメータのqwen3.8-maxは、数日間にわたり自律的にソフトウェアを構築し研究論文を再現。長期稼働型エージェント(long-horizon agents)へのシフトを示唆している。
alibaba unveiled qwen3.8-max, a 2.4-trillion-parameter model designed to handle complex, multi-day tasks without human intervention. in internal tests, the model autonomously built a command-line tool with 265 commits over 16 days, reproduced and improved upon research paper results, and outperformed 458 human teams in a coding challenge. the model uses a mixture-of-experts architecture with 95b active parameters per query, matching the performance of top western proprietary models. by demonstrating sustained autonomous reasoning and planning, qwen3.8-max sets a new standard for open-weight agents and intensifies the race for agentic systems. weights will be released publicly next week.