DeepSeek V3.2

DeepSeek V3.2 is DeepSeek’s update to V3.1, publicly detailed on December 1, 2025 after an experimental preview in September. It introduces DeepSeek Sparse Attention, a fine-grained sparse attention mechanism that cuts training and inference cost while preserving quality on long-context tasks, and the flagship Speciale variant carries roughly 685 billion parameters in a mixture-of-experts layout. The context window sits at 163,840 tokens with up to 64,000 tokens of output, and DeepSeek reports gold-medal-level results on 2025 International Mathematical Olympiad and International Olympiad in Informatics problem sets. API pricing runs around $0.21 per million input tokens and $0.32 per million output tokens, roughly half of what V3.1-Terminus cost, particularly for long-context decoding. Like the rest of the DeepSeek lineup it ships under the MIT license with weights on Hugging Face, and it served as the base DeepSeek built V4-Pro on the following spring.