Qwen3 Coder Next

Qwen3 Coder Next is a specialized language model engineered to serve as a robust foundation for coding agents and local development environments. Built upon a hybrid architecture that integrates gated attention and a Mixture of Experts design, the model achieves a total of 80 billion parameters while activating only 3 billion during inference. This design intent focuses on delivering high-level reasoning and coding capabilities within a lightweight active footprint, allowing it to maintain performance comparable to much larger models while significantly reducing the hardware requirements for deployment.

The model benefits from an extensive training lineage that emphasizes agentic capabilities, including supervised fine-tuning on high-quality agent trajectories and continued pretraining on code-centric data. By incorporating large-scale executable task synthesis and reinforcement learning from environment feedback, the model is specifically optimized for long-horizon reasoning, complex tool usage, and reliable recovery from execution failures. These advancements make it a versatile tool for developers, offering seamless integration with various IDEs and CLI platforms to support dynamic, real-world coding workflows.