DeepSeek-V3.2 unveils DSA sparse-attention, expert distillation and 1,800-task simulators

DeepSeek-V3.2 technical report presents three core innovations: DeepSeek Sparse Attention (DSA) that focuses on key text fragments and reduces computation growth to linear; a training pipeline of specialist expert models plus large-scale RL knowledge distillation; and synthetic simulators of 1,800 tasks for tool-use RL. The report states IMO/IOI gold-level results, parity with Google’s Gemini-3.0-Pro on reasoning benchmarks, and superiority to GPT-5 on some complex tasks. Paper: https://t.co/pveCm8ZVoK