On May 8, Alibaba's Qwen team released Qwen3-Coder-Next, an 80B-parameter MoE model purpose-built for Agentic Coding, activating only 3B parameters per forward pass yet possessing programming capability comparable to Dense models, sparking wide discussion in the open-weights community.
The core technology lies in Gated DeltaNet combined with Gated Attention, converting Attention's quadratic compute complexity to linear, allowing the model to maintain 262K token ultra-long context while avoiding exponential-order latency penalties. On repository-level tasks, throughput improves about 10× compared to same-tier Dense models. The training stage introduces a Best-Fit Packing strategy, effectively mitigating hallucination issues in long-context scenarios while preserving the integrity of contextual information.
The model is open-sourced under the Apache 2.0 license, with weights published in 4 variants on HuggingFace, along with a detailed technical report. At a time when coding agents are becoming the competitive focus of the industry, the small-activation-large-capability sparse MoE design provides a new paradigm for local deployment — developers can obtain 80B-tier structured code understanding capability at the resource cost of a 3B model on consumer-grade hardware, directly challenging the capability ceiling previously only accessible to closed-source large models.