Persona: Senior AI Research Lead & Systems Architect
Abstract:
This technical report details the release of Qwen3-Coder-Next, an open-weight language model optimized for agentic coding and local development environments. Built upon a novel hybrid attention and Mixture-of-Experts (MoE) architecture (80B total / 3B active parameters), the model prioritizes agentic training signals over pure parameter scaling. By utilizing executable task synthesis, environment-feedback loops, and expert distillation, the model achieves high performance on complex benchmarks such as SWE-Bench Pro while maintaining a highly efficient Pareto frontier. The report concludes that scaling agentic turns and multi-turn reasoning is more critical for real-world software engineering tasks than traditional scaling laws.
Summary of Qwen3-Coder-Next Technical Report:
[0:00] Model Architecture: Qwen3-Coder-Next is built on the Qwen3-Next-80B-A3B-Base foundation, utilizing a hybrid attention mechanism and a Mixture-of-Experts (MoE) structure with 3 billion active parameters.
[0:30] Agentic Training Scaling: The training methodology shifts focus from parameter count to the quality of agentic signals. This involves continued pretraining on code-centric data, supervised fine-tuning on high-quality agent trajectories, and domain-specialized expert training in software engineering and UX.
[0:45] Verifiable Environments: The model is trained using large-scale executable task synthesis, allowing it to learn directly from environment feedback, long-horizon reasoning, and autonomous recovery from execution failures.
[1:00] SWE-Bench Verified Performance: Utilizing the SWE-Agent scaffold, the model achieves a success rate exceeding 70% on SWE-Bench Verified. It remains competitive in multilingual environments and the more rigorous SWE-Bench Pro category.
[1:15] Long-Horizon Reasoning: Performance on SWE-Bench Pro improves significantly as the number of agent turns increases, demonstrating the model's capacity for complex, multi-turn problem-solving.
[1:30] Efficiency Pareto Tradeoff: The model occupies a unique position on the efficiency frontier; with only 3B active parameters, it matches or exceeds the performance of open-source models with 10× to 20× the active parameter count.
[1:45] Practical Applications & Demos: The report highlights successful integration into various downstream applications, including:
Web Development: Automated generation of chat interfaces and site components.
CLI & Desktop: Autonomous terminal usage and desktop environment management.
Agentic Tools: Compatibility with OpenClaw, Claude Code, Cline, and browser-use agents for product searching and game building.
[2:00] Strategic Conclusion: Future development will focus on enhancing autonomous tool usage, complex task management, and iterative updates based on real-world user interaction data.