Qwen3-Coder-Next Released, Reducing Inference Costs via Agent-Specific Training

A key characteristic of this model is its focus on scaling agentic training signals rather than relying solely on parameter scaling. It was trained using a combination of verifiable coding tasks and executable environments, directly incorporating feedback from those environments. This has strengthened its capabilities in long-term reasoning, tool use, and recovery from execution failures.

In benchmarks such as SWE-Bench Pro, the model has achieved positive results by increasing the number of agent turns. By improving the trade-off between efficiency and performance, it achieves high coding proficiency with lower inference costs. Alibaba plans to further improve agent skills, such as tool use and complex task management, in the future.