Qwen 3.5 35B A3B

• 🆕 Sparse MoE with 35B total parameters, only 3B active
• 🏢 Built by Alibaba's Qwen team, Apache 2.0 licensed
• 📏 Native 262K-token context per the model card
• 👁️ Multimodal inputs spanning text, vision, and code
• 🎯 Targets reasoning, coding, and agentic workflows
• 🔧 Function-calling support for agentic integrations
• 🌐 Built-in tool-use and web-search capabilities
• 📚 Part of Alibaba's Qwen 3.5 model generation

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen 3.5 35B A3B is an efficiency-focused entry in Alibaba's Qwen 3.5 lineup, released in February 2026. It uses a sparse Mixture-of-Experts design that holds 35 billion total parameters but activates just 3 billion per token, keeping inference cheap while retaining broad capacity. The model card lists a native 262K-token context window, and the weights are distributed under the permissive Apache 2.0 license.

Within the 3.5 generation, it sits above the dense Qwen 3.5 9B and below the heavyweight Qwen 3.5 397B and Qwen3.5 122B A10B siblings, offering a middle option that trades raw scale for low active-compute cost. It also follows earlier Qwen flagships such as Qwen 3 235B A22B Instruct 2507 in the broader family.

Functionally, the model targets reasoning, coding, and agentic workflows, with vision, function-calling, and web-search capabilities exposed through the catalog. Its successor, Qwen 3.6 27B, continues the same MoE family with further refinements. For developers, the practical appeal is pairing a long-context, multimodal model with a small active-parameter footprint.