We present Qwen3-Coder-Next, an open-weight language model specialized for coding agents. Qwen3-Coder-Next is an 80-billion-parameter model that activates only 3 billion parameters during inference, enabling strong coding capability with efficient inference. In this work, we explore how far strong training recipes can push the capability limits of models with small parameter footprints. To achieve this, we perform agentic training through large-scale synthesis of verifiable coding tasks paired with executable environments, allowing learning directly from environment feedback via mid-training and reinforcement learning. Across agent-centric benchmarks including SWE-Bench and Terminal-Bench, Qwen3-Coder-Next achieves competitive performance relative to its active parameter count. We release both base and instruction-tuned open-weight versions to support research and real-world coding agent development.
核心贡献 · Key contributions
提出 Qwen3-Coder-Next，一种 80B MoE 模型，仅 3B 活跃参数，专为编码智能体设计。 Proposes Qwen3-Coder-Next, an 80B MoE model with 3B active parameters for coding agents.
开发了大规模智能体训练，包含可验证任务和可执行环境。 Develops large-scale agentic training with verifiable tasks and executable environments.
在 SWE-Bench 上取得了与活跃参数数量相称的竞争性表现。 Achieves competitive SWE-Bench performance relative to active parameter count.
引入多模板工具调用训练，实现鲁棒的格式泛化。 Introduces multi-template tool-call training for robust format generalization.
在竞争性编程之外的多样化编码任务中应用执行驱动的强化学习。 Applies execution-driven RL across diverse coding tasks beyond competitive programming.
发布基础版和指令微调版开源权重，支持研究与开发。 Releases base and instruction-tuned open-weight versions for research and development.
局限 · Limitations
活跃算力和总训练算力均小于前沿专有模型。 Smaller active compute and total training compute than frontier proprietary models.
在解决高度复杂的大规模软件工程任务方面存在差距。 Gap in solving highly complex, large-scale software engineering tasks.
复杂任务可能需要更多交互轮次，影响效率。 May require more interaction turns for complex tasks, affecting efficiency.
前端和 UI 相关能力仍有改进空间。 Frontend and UI-related capabilities remain an area for improvement.
智能体训练轨迹中观察到跨框架迁移有限。 Limited cross-scaffold transfer observed in agentic training trajectories.
论文章节 · Sections（共 15）
摘要Abstract
1 引言1 Introduction
2 扩展智能体训练2 Scaling up Agentic Training
2.1 任务合成2.1 Task Synthesis
2.2 基础设施2.2 Infrastructure
3 中期训练3 Mid-training
3.1 数据3.1 Data
3.2 训练3.2 Training
4.1 监督微调4.1 Supervised Fine-tuning
4.2 专家模型4.2 Expert Models
5.1 智能体评估5.1 Agentic Evaluation
5.2 其他编码任务5.2 Other Coding Tasks
5.3 通用任务5.3 General Tasks
6 结论、局限与未来工作6 Conclusion, Limitation, and Future Work