DeepSeek‑V3.2 introduces DSA sparse attention and efficient training, matches top closed models

DeepSeek‑V3.2 introduces a DeepSeek Sparse Attention (DSA) mechanism that reduces long‑text computation from exponential to linear while preserving understanding. Training used specialist expert models, knowledge distillation and large‑scale RL to transfer domain reasoning. Synthetic simulation of 1,800+ virtual tasks improved tool use and instruction following. Benchmarks reported IMO/IOI gold performance and parity with Gemini‑3.0‑Pro, outperforming GPT‑5 on some tasks. Paper: https://t.co/pveCm8ZVoK