DeepResearch Agent System achieves 3.2x faster inference via sparse activation

The DeepResearch Agent System uses a 30B parameter sparse architecture with 3B active parameters per token to accelerate research tasks. It features a dual-mode reasoning engine and hierarchical attention, improving recall by 23.4% over standard long-context methods.