Qwen3-Coder-Next is an open-weight language model from Alibaba, specifically engineered for agentic coding and local software development. It features a hybrid architecture that combines Gated DeltaNet (linear attention) with Mixture-of-Experts (MoE), facilitating high throughput and efficient long-context modeling. The model is designed to function as a backbone for AI coding agents, supporting a native context length of 256K tokens.
Despite its total count of 80 billion parameters, Qwen3-Coder-Next only activates 3 billion parameters per token. This sparse MoE design allows the model to deliver performance on benchmarks like SWE-Bench Verified that matches or exceeds significantly larger dense models. Its training focuses on scaling agentic signals, utilizing verifiable coding tasks paired with executable environments to improve long-horizon reasoning and tool usage.
The model is optimized for real-world development tasks, including managing complex workflows, autonomous debugging, and recovering from execution errors. Unlike some reasoning-heavy variants in the Qwen series, Qwen3-Coder-Next is a non-thinking model, meaning it does not generate intermediate <think> blocks. It is designed for seamless integration with various command-line and IDE-based agent frameworks.