Qwen3.5
Open-sourced a 397B-parameter MoE model that activates only 17B parameters per inference.
397B / 17Btotal / active parameters
Why it matters
Hybrid linear attention makes long-context inference cheaper, and Chinese open models stayed in the top tier.
Built on
Source: qwen.ai ↗
Official page not opened for verification; URL taken from citations in multiple reports.