Alibaba Qwen Team Releases Qwen3-Coder-Next: Open-Source Ultra-Sparse Model Boosts Repository-Level Code Task Throughput by 10x
Alibaba’s Qwen team recently introduced Qwen3-Coder-Next, an open-source, weight-shared language model specifically designed for code agents and local development. Based on an ultra-sparse Mixture-of-Experts architecture, the model achieves a 10x throughput increase in repository-level code tasks while supporting ultra-long context processing of 262,144 tokens. The model weights are openly available on the Hugging Face platform under the Apache 2.0 license.
Model Architecture and Core Technologies
Qwen3-Coder-Next employs an ultra-sparse MoE architecture with 80 billion total parameters, activating only 3 billion parameters per forward pass. This design, combined with Gated DeltaNet and Gated Attention technologies, effectively addresses the quadratic complexity issue of traditional Transformers in long contexts. The model also introduces a Best-Fit Packing mechanism to reduce context-related hallucinations and supports a tool-calling format in an XML style.
Performance and Benchmarks
In repository-level tasks, Qwen3-Coder-Next achieves a 10x throughput increase compared to dense models of similar capacity. It can read entire Python libraries or complex JavaScript frameworks and covers 370 programming languages. It scored 70.6% on the SWE-Bench Verified benchmark, outperforming DeepSeek-V3.2, and 61.2% in the SecCodeBench code generation scenario, surpassing Claude-Opus-4.5. While maintaining the advantage of lightweight local deployment, the model demonstrates reasoning capabilities comparable to large-scale proprietary systems.
Training Methodology and Data Sources
The model’s training utilizes the “closed-loop” agent training workflow, MegaFlow, processing 800,000 verifiable coding tasks from GitHub pull requests on Alibaba Cloud’s native scheduling system. The Qwen team stated that scaling up agent training, rather than just increasing model size, is the key driver for enhancing real-world coding agent capabilities. Additionally, the model integrates specialized expert models for web development and user experience, which are distilled into the main model.
Open-Source Release and Application Prospects
Qwen3-Coder-Next is available in four variants, with weights uploaded to Hugging Face for easy download and deployment by developers. The model is optimized for cross-file dependency logic, making it suitable for scenarios like intelligent coding, unit testing, and legacy system migration. It provides the open-source community with an efficient, localized code agent solution.
