Qwen3-Coder-Next Technical Report
Ruisheng Cao, Mouxiang Chen, Jiawei Chen, Zeyu Cui, Yunlong Feng, Binyuan Hui, Yuheng Jing, Kaixin Li, Mingze Li, Junyang Lin, Zeyao Ma, Kashun Shum, Xuwu Wang, Jinxi Wei, Jiaxi Yang, Jiajun Zhang, Lei Zhang, Zongmeng Zhang, Wenting Zhao, Fan Zhou
2026-03-04
Summary
This paper introduces Qwen3-Coder-Next, a new language model designed specifically for writing and understanding code, and explores how to make smaller models perform really well at coding tasks.
What's the problem?
Creating AI that can reliably write code is challenging. Larger models generally perform better, but they require a lot of computing power to run, making them expensive and slow. The core issue is how to get strong coding ability from a model that doesn't have a huge number of parameters, which would make it more practical to use.
What's the solution?
The researchers built an 80-billion parameter model, but cleverly designed it to only *use* 3 billion parameters at a time when actually generating code. They then trained this model using a special method where it learns by trying to solve coding problems in simulated environments and getting feedback on whether its code works. This training included both learning from the results of its code and using reinforcement learning to improve over time.
Why it matters?
This work is important because it shows that you can achieve impressive coding performance with a relatively small and efficient model. By releasing the model to the public, the researchers are helping other developers and researchers build better coding tools and AI assistants without needing massive amounts of computing resources. It pushes the boundaries of what's possible with smaller AI models, making them more accessible and practical for real-world applications.
Abstract
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.