Qwen3-Coder-Next 是阿里云通义千问团队推出的开源代码 Agent 专用模型,其核心优势在于通过 MoE 架构(80B 总参数,3B 激活)实现了极高的效率与性能平衡,并通过大规模代理式训练(Agentic Training)在复杂 Agent 任务上取得了突破。
| Benchmark | Qwen3-Coder-Next (80A3) | Claude-Sonnet-4.5 (?) | DeepSeek-V3.2 (671A37) |
|---|---|---|---|
| SWE-Bench Verified (SWE-Agent) | 70.6% | 76.0% | 70.2% |
| SWE-Bench Pro (SWE-Agent) | 44.3% | 44.3% | 40.9% |
| Terminal-Bench 2.0 (Terminus2-json) | 36.2% | 51.7% | 39.3% |
| Codeforces (Rating) | 2100 | 1875 (Qwen3-Next) | 1800 (Qwen3-Coder-480) |