Qwen3-Coder and Qwen3.5-Plus - Agentic Coding in the World!
如果无法正常显示,请先停止浏览器的去广告插件。
1. Qwen3-Coder and Qwen3.5-Plus
Agentic Coding in the World!
崔泽宇
Qwen Team, Alibaba Group
2. Why Coding is Important
Verifier 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://www.jasonwei.net/blog/asymmetry-of-verification-and-verifiers-law
3. 我们在持续的拥抱Coding 与 开源
Foundation Models
2024.04
CodeQwen1.5
Qwen
2024.11
Qwen2.5
Coding Copilot (FIM)
Qwen3-Coder-
Family
2025.07.21💪 Qwen3-Coder-Plus: Qwen3-Coder-480B-A35B
2025.07.31⚡ Qwen3-Coder-Flash: Qwen3-Coder-30B-A3B
2026.02.02🔧 Qwen3-Coder-Next: Qwen3-Coder-80B-A3B
最快并且强的agentic coding 模型
Qwen3.5-Plus: Qwen3.5-397B-A17B
第
次图
融合的、全领域的模型
Qwen2.5-Coder
Coding Assistant
2025.07
Qwen3
Qwen3-Coder
Coding Agent
4. Qwen3-Coder-Next:Strong and Fast
Qwen3-Coder-Next: 接近 Qwen3-Coder-Plus 的效果、超越Qwen3-Coder-Flash的速
度
混合MoE 架构、 规模可执 任务合成、格式扩展、强化学习
•Qwen3-Coder-Next(3B 激活) 表现可与 激活参数量
•Qwen3-Coder-Next 在
向低成本代码智能体部署
10×–20× 的模型相当。
具有较为明显的优势。
5. Qwen3.5-Plus:迈向原 多模态智能体
Qwen3.5-Plus:也被称作 Qwen3.5-397B-A17B
Coding Agent 的best practice
复杂任务:Qwen3.5-Plus
简单任务:Qwen3-Coder-Next
6. Case
HTML/JS 农场模拟游戏(星露
常AI模型发布情况
总结
格)
7. How we train Qwen-Coder?
Scaling Law: “
就是好”
• Scaling Data for a Stronger Foundation
• Scaling Formats, Improving Instruct Following
• Scaling Tasks for Greater Diversity
8. 1 Scaling Data for a Stronger Foundation
Pre-training: Qwen3-Coder-Plus Scaling long horizon data make great progress
⭐ Key Highlights
Scaling Tokens: 31T (base) + 7.5T (Code mix,
70% code ratio), excelling in coding while
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 quality
9. 1 Scaling Data Scaling Foundation
Pre-training: Scaling high quality data
Coder
Coder
Data
Data
Coder
Data
10. 2 Scaling Format, Scaling Instruct Following
各家公司都有
https://arxiv.org/pdf/2603.00729
的Toolcall格式定义规范
11. 2 Scaling Format, Scaling Instruct Following
12. 3 Scaling Tasks for Greater Diversity
Post-training:
Virtualized Sandbox
Agent Service Kit
13. 3 Scaling Tasks for Greater Diversity
Post-training: Scaling tasks and Scaling Envs
SWE-Universe: 从 Github 海量的Pull Request
动构造单测、构造环境、校验PR可执
形成了80W 条 训练实例 (执
https://arxiv.org/pdf/2602.02361
性
环境 + 任务描述 + 真实patch + 可验证 脚本)
14. 3 Scaling Tasks for Greater Diversity
Post-training: Scaling Envs and Scaling tasks
限多个可执
理论上可以扩充
动bug 引
: 通过
从上述 80W 条 训练实例
验证的训练样本
15. Coding Moment
16. 3 Scaling Tasks for Greater Diversity
17. 欢迎讨论
18. Thanks!