📄 中文摘要
Qwen团队近日发布了Qwen3-Coder-Next，这是一款专为编码代理（Coding Agents）和本地开发设计的开源权重语言模型。该模型基于Qwen3-Next-80B-A3B骨干网络构建，采用稀疏专家混合（Sparse Mixture-of-Experts, MoE）架构，并结合混合注意力机制（Hybrid Attention）。其总参数规模达800亿，但每个token仅激活30亿参数，这种高效设计显著降低了推理计算成本，使其能够在消费级硬件上实现高性能运行，特别适合本地开发环境。
技术要点上，Sparse MoE架构是核心创新之一。
📄 English Summary
Qwen Team Releases Qwen3-Coder-Next: An Open-Weight Language Model Designed Specifically for Coding Agents and Local Development
[Qwen Team Releases Qwen3-Coder-Next: An Open-Weight Language Model Designed Specifically for Coding Agents and Local Development](https://www.marktechpost.com/2026/02/03/qwen-team-releases-qwen3-coder-next-an-open-weight-language-model-designed-specifically-for-coding-agents-and-local-development/) The Qwen team has unveiled Qwen3-Coder-Next, an open-weight large language model (LLM) meticulously engineered for coding agents and local development workflows. Built upon the robust Qwen3-Next-80B-A3B backbone, it employs a sparse Mixture-of-Experts (MoE) architecture augmented with hybrid attention mechanisms. Boasting 80 billion total parameters yet activating only 3 billion per token, this design achieves remarkable inference efficiency, enabling deployment on consumer-grade hardware without compromising performance. At its core, the sparse MoE framework dynamically routes tokens to a subset of specialized expert modules, minimizing computational overhead while preserving the model's expansive knowledge capacity. This contrasts with dense models where all parameters are engaged per inference step, often leading to prohibitive latency and memory demands. The hybrid attention mechanism innovatively blends local attention for fine-grained code token dependencies with global attention for broader contextual understanding, excelling in long-sequence tasks like multi-file codebases or repository-level analysis. Inherited from the Qwen3-Next series, enhancements include superior multilingual code support, extended context windows (up to 128K tokens), and task-specific fine-tuning on vast coding corpora encompassing 20+ programming languages. Key innovations lie in its agent-centric optimization. Unlike general-purpose LLMs, Qwen3-Coder-Next is pre-trained and aligned for autonomous coding agents via agent-oriented reinforcement learning in simulated dev environments. It natively supports tool-use (e.g., shell execution, Git APIs), multi-step planning, self-reflection, and iterative refinement—critical for agentic workflows.