Agentic Coding构建之路：从模型到框架-崔泽宇.pdf

1、Qwen3-Coder and Qwen3.5-Plus Agentic Coding in the World!Qwen Team,Alibaba Group崔泽宇Why Coding is ImportantVerifier Law：训练 AI 解决某个任务的容易程度，与该任务的可验证性成正。Coding 是绝佳的可验证场景。https:/ 与 开源CodeQwen1.5Qwen2.5-CoderQwen3-Coder2024.042024.112025.07QwenQwen2.5Qwen3Foundation ModelsCoding Copilot(FIM)Coding Assista

2、ntCoding AgentQwen3-Coder-Family Qwen3-Coder-Plus：Qwen3-Coder-480B-A35B Qwen3-Coder-Flash：Qwen3-Coder-30B-A3B Qwen3-Coder-Next：Qwen3-Coder-80B-A3B 2025.07.212025.07.312026.02.02Qwen3.5-Plus：Qwen3.5-397B-A17B 最快并且强的agentic coding 模型 第次图融合的、全领域的模型Qwen3-Coder-Next：Strong and Fast混合MoE 架构、规模可执任务合成、格式扩展、

3、强化学习Qwen3-Coder-Next：接近 Qwen3-Coder-Plus 的效果、超越Qwen3-Coder-Flash的速度Qwen3-Coder-Next（3B 激活）表现可与激活参数量 1020的模型相当。Qwen3-Coder-Next 在向低成本代码智能体部署具有较为明显的优势。Qwen3.5-Plus：迈向原多模态智能体Coding Agent 的best practice 复杂任务：Qwen3.5-Plus 简单任务：Qwen3-Coder-Next Qwen3.5-Plus：也被称作 Qwen3.5-397B-A17BCaseHTML/JS 农场模拟游戏（星露格）总结常

4、AI模型发布情况How we train Qwen-Coder?Scaling Data for a Stronger Foundation Scaling Formats,Improving Instruct Following Scaling Tasks for Greater Diversity Scaling Law：“就是好”1 Scaling Data for a Stronger Foundation Key Highlights Scaling Tokens:31T(base)+7.5T(Code mix,70%code ratio),excelling in coding w

5、hile preserving general and math abilities Scaling Context:Natively supports long context(256K)with focus on repo-scale and dynamic data(e.g.,Pull Requests)for Agentic Coding High-Quality Synthetic Data:Leveraged Qwen2.5-Coder to clean and rewrite noisy data,significantly improving overall data qual

6、ityPre-training:Qwen3-Coder-Plus Scaling long horizon data make great progress 1 Scaling Data Scaling FoundationDataCoderDataCoderDataCoderPre-training:Scaling high quality data2 Scaling Format,Scaling Instruct Following各家公司都有的Toolcall格式定义规范https:/arxiv.org/pd