Qwen3-Coder-Next: The Evolution of a Small Hybrid Model Aimed at Agentic Coding

Qwen3-Coder-Next is an open-weight model built for coding agents and local development, constructed on top of Qwen3-Next-80B-A3B-Base. Building on a foundation that uses hybrid attention and MoE, it scales up agentic training centered on large-scale executable tasks, environment interaction, and reinforcement learning.

The key idea is scaling agentic training signals rather than scaling parameters. The model learned to use tools, reason over long horizons, and recover from execution failures like a real agent, by training on a large volume of verifiable coding tasks paired with execution environments.

The training composition includes:

• Continued pretraining on code- and agent-centric data
• Supervised fine-tuning on data containing high-quality agent trajectories
• Domain-specific expert training in areas like software engineering, QA, and web/UX
• Expert distillation that merges multiple experts into a single model for deployment

The results show up in coding agent benchmarks. On SWE-Bench Verified, it surpassed 70% using the SWE-Agent scaffold, and it also showed competitiveness in multilingual settings and on the more difficult SWE-Bench Pro. Notably, with a small active parameter scale of just 3B active, it achieved agent-centric results comparable to or better than some much larger open-source models.

On SWE-Bench Pro, performance also improved as the number of agent turns increased, demonstrating long-horizon reasoning capability in multi-turn agentic tasks. In terms of the balance between efficiency and performance, it achieves SWE-Bench-Pro performance comparable to models with 10x to 20x larger active parameters, placing it on a strong Pareto frontier for cost-effective agent deployment.

Demos were presented across various downstream environments including OpenClaw, Qwen Code, Claude Code, web dev, browser use, and Cline. Through examples such as building web interfaces, creating games, ASCII art tools, desktop organization, browser search, and website testing, it emphasizes that a small model can be fully capable of fitting into real agent workflows.

Looking ahead, the direction includes further strengthening reasoning and decision-making, supporting more tasks, and rapidly updating in line with user usage patterns. In summary, this model pushes the idea that agentic capabilities—tool use, long-horizon reasoning, and execution recovery—rather than simple scale competition, are core to the next generation of coding agents.