Qwen3-Coder-Next

Qwen3-Coder-Next is a specialized language model engineered to serve as a high-performance engine for coding agents and local development environments. Built upon a hybrid architecture that integrates Mixture-of-Experts with gated attention and DeltaNet layers, the model manages 80 billion total parameters while activating only 3 billion during each inference pass. This design intent focuses on delivering elite coding capabilities—such as long-horizon reasoning and complex tool usage—within a compact active footprint, allowing developers to run sophisticated agentic workflows on consumer-grade hardware.

The model’s development lineage centers on a rigorous agentic training recipe that emphasizes learning from environment feedback. By utilizing large-scale synthesis of verifiable coding tasks and executable environments, the model underwent continued pre-training and supervised fine-tuning on high-quality agent trajectories. This approach enables the model to excel at recovering from execution failures and navigating dynamic coding tasks. With its support for extensive context lengths and compatibility with various CLI and IDE platforms, it is positioned as a versatile tool for both individual developers and enterprise-scale coding agent deployments.