Qwen3-Coder-Next：推进小型混合模型在智能体编程中的应用

Tech ReportGitHubHugging FaceModelScopeDISCORD

Introduction#

We introduce Qwen3-Coder-Next, an open-weight language model designed specifically for coding agents and local development. Built on top of Qwen3-Next-80B-A3B-Base, which adopts a novel architecture with hybrid attention and MoE, Qwen3-Coder-Next has been agentically trained at scale on large-scale executable task synthesis, environment interaction, and reinforcement learning, obtaining strong coding and agentic capabilities with significantly lower inference costs.

Scaling Agentic Training#

Rather than relying solely on parameter scaling, Qwen3-Coder-Next focuses on scaling agentic training signals. We train the model using large collections of verifiable coding tasks paired with executable environments, enabling the model to learn directly from environment feedback. This includes:

• Continued pretraining on code- and agent-centric data
• Supervised fine-tuning on data containing high-quality agent trajectories
• Domain-specialized expert training (e.g., software engineering, QA, web/UX)
• Expert distillation into a single deployment-ready model

This recipe emphasizes long-horizon reasoning, tool usage, and recovery from execution failures, which are essential for real-world coding agents.

Performance on Coding Agent Benchmarks#

Agent-Centric Benchmark Results#

The figure below summarizes performance across several widely used coding agent benchmarks, including SWE-Bench (Verified, Multilingual, and Pro), TerminalBench 2.0, and Aider.

The figure demonstrates that:

• Qwen3-Coder-Next achieves over 70% on SWE-Bench Verified using the SWE-Agent scaffold.
• Performance remains competitive across multilingual settings and the more challenging SWE-Bench Pro benchmark.
• Despite its small active footprint, the model matches or exceeds several much larger open-source models across agent-centric evaluations.

As shown in the figure below, Our model achieves strong results on SWE-Bench Pro by scaling the number of agent turns, providing evidence that the model excels at long-horizon reasoning in multi-turn agentic tasks.

Efficiency–Performance Tradeoff#

This figure highlights how Qwen3-Coder-Next achieves an improved Pareto tradeoff between efficiency and performance.

This comparison makes the efficiency story clear:

• Qwen3-Coder-Next (3B active) achieves SWE-Bench-Pro performance comparable to models with 10×–20× more active parameters.
• Qwen3-Coder-Next sits on a strong Pareto frontier for cost-effective agent deployment.

Demo#

The small and fast coder model can be integrated into different downstream applications, and below we demonstrate showcases of OpenClaw, Qwen Code, Claude Code, Web dev, browser use, Cline, etc.

Web Dev#

Demo1 Creating a Chat Interface

CLI#

Demo1 Desktop Cleanup

Cline#

Demo1 Creating a Multicolor Animation

OpenClaw#

Demo1 Web Page for Qwen3 Coder Next

Browser Use Agent#

Demo1 Searching for a Product on Amazon

coder.qwen.ai#

Demo1 Building a Gomoku Game

Summary and Future Work#

Qwen3-Coder-Next shows promising results on coding agent benchmarks, providing good speed and reasoning abilities for practical use. While it performs competitively—even compared to some larger open-source models—there is still much room for improvement.

Looking ahead, we believe strong agent skills—like using tools by itself, handling tough problems, and managing complex tasks—are key for better coding agents. Next, we plan to improve the model’s reasoning and decision-making, support more tasks, and update quickly based on how people use it.

Citation#

If you find our work helpful feel free to give us a cite.

@techreport{qwen_qwen3_coder_next_tech_report,
  title        = {Qwen3-Coder-Next Technical Report},
  author       = {{Qwen Team}},
  url          = {https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf},
  note         = {Accessed: 2026-02-03}
}