Alibaba Qianwen Launches Open Source Weighted Language Model Qwen3-Coder-Next

On February 4th, Ali Thousand Qwen introduced Qwen3-Coder-Next, an open-source weighted language model designed specifically for programming intelligence and local development. The model is based on Qwen3-Next-80B-A3B-Base and adopts a new architecture of hybrid attention and MoE. By training the intelligent agent through large-scale executable task synthesis, environment interaction, and reinforcement learning, it achieves powerful programming and intelligent agent capabilities while significantly reducing inference costs. According to the introduction, Qwen3-Coder-Next does not rely solely on parameter expansion, but focuses on expanding the intelligent agent training signal. Ali Thousand Qwen uses large-scale verifiable programming tasks and executable environments for training, allowing the model to learn directly from environment feedback. The training process includes continuous pre-training on data centered around code and intelligent agents, supervised fine-tuning on data containing high-quality intelligent agent trajectories, domain-specific expert training (such as software engineering, QA, Web/UX, etc.), and distilling expert capabilities into a single, deployable model. This recipe emphasizes long-term reasoning, tool usage, and recovery from execution failures, which are crucial for programming intelligent agents in the real world.