Alibaba Qwen Open-Sources Qwen3.6-35B-A3B: Lightweight & Efficient MoE Model with Excellent Agent Programming Capabilities

Alibaba Qwen large model today announced the open-source release of Qwen3.6-35B-A3B, a sparse Mixture of Experts (MoE) model with 35 billion total parameters and only 3 billion active parameters. The model is lightweight and efficient, with outstanding agent programming capabilities—it significantly outperforms its predecessor and competes with dense models of the same scale. It also supports multi-modal thinking and non-thinking modes, making it one of the most versatile open-source models available.

Currently, Qwen3.6-35B-A3B is live on Qwen Studio, with open-source weights on Hugging Face and ModelScope. Users can call it via Alibaba Cloud Bailian API as qwen3.6-flash, and it integrates seamlessly into third-party programming assistants like OpenClaw and Claude Code to simplify development workflows.