Alibaba Open-Sources Qwen3.6-35B-A3B with Just 3B Active Parameters, Targeting Top-Tier Agent Coding

TL;DR AI
2 min readKey summary
Alibaba’s Qwen team released Qwen3.6-35B-A3B, an open-source MoE model.
Although it has 35B parameters overall, it activates only 3B per request to improve compute efficiency.
The model is built for strong agent coding performance and also supports multimodal and non-reasoning use cases.
The release gives developers a lower-cost open model with competitive advanced AI capabilities.



