DeepSeek released two new models, DeepSeek-V3.2 and high-compute variant DeepSeek-V3.2-Speciale.

A major technical breakthrough of the new models is the DeepSeek Sparse Attention mechanism, which slashes computational costs for processing long documents while maintaining model performance, DeepSeek said in its technical report.

DeepSeek-V3.2 performs comparably to GPT-5, while DeepSeek-V3.2-Speciale surpasses GPT-5 and exhibits reasoning proficiency comparable to Gemini-3.0-Pro, according to the report.