Alibaba Unveils Qwen 3.5: Gated Delta Networks and Sparse MoE for Agentic AI

Alibaba Group has officially introduced Qwen 3.5, its next-generation artificial intelligence platform. According to the company, this release marks a significant technical pivot toward "agentic AI," centering on a model engineered to perform complex, autonomous operations rather than simple conversational tasks.

​New Architecture: Gated Delta Networks + MoE

​At the core of Qwen 3.5 is a sophisticated new architecture that, according to Alibaba, merges Gated Delta Networks with a sparse mixture-of-experts (MoE) framework. This design addresses the trade-off between model size and inference speed:

• ​Active vs. Total Parameters: Alibaba states that while the model boasts a massive total count of 397 billion parameters, the sparse MoE structure allows it to activate only 17 billion parameters during each forward pass.
• ​Efficiency: According to the developers, this selective activation significantly reduces computational load, optimizing the model for high-speed token generation and deployment economics.

​Agentic Capabilities and Long Context

​Qwen 3.5 is designed to function as an autonomous agent capable of planning and executing multi-step workflows.

• ​Visual Agents: Alibaba highlights the model’s advanced visual agentic capabilities, which allow it to perceive software interfaces, make independent decisions, and execute actions across various applications.
• ​1 Million Tokens: The company notes that the closed-source iteration, Qwen-3.5-Plus, offers a 1-million-token context window. This capacity facilitates deep document analysis and long-form reasoning without losing coherence.

​Global Reach and Open Source

​To support global deployment, Alibaba reports that it has significantly broadened the model's linguistic reach. Qwen 3.5 now supports 201 distinct languages and dialects (up from 119 in the previous version), which the company claims strengthens its performance in Southeast Asian and European markets.

Continuing its open-weights strategy, Alibaba has released Qwen-3.5-Open, providing the research community with access to these advanced architectural improvements.

​Performance Gains

​According to Alibaba, the architectural optimizations in Qwen 3.5 have resulted in an eightfold improvement in handling large workloads compared to the Qwen 3 series. In self-reported benchmarks, the company asserts that the model delivers performance commensurate with leading global systems, specifically citing parity with GPT-5.2 and Claude Opus 4.6 in reasoning and coding tasks.