DeepSeek has launched its V3.2-Exp experimental model featuring a revolutionary sparse attention mechanism that dramatically reduces computational costs while maintaining performance parity with previous models. Released on September 29, 2025, this breakthrough represents a significant advancement in efficient large language model architecture.
The key innovation driving these improvements is the DeepSeek Sparse Attention (DSA) mechanism, which fundamentally changes how the model processes information. Instead of computing attention between every token pair in the traditional O(n²) complexity approach, DSA selectively processes approximately 30% of token pairs, achieving computational complexity closer to O(n log n).
Performance benchmarks demonstrate that V3.2-Exp maintains the high reasoning capabilities of its predecessor, achieving an MMLU-Pro score of 85.0 – identical to V3.1-Terminus. The model shows improved performance in coding tasks, with its Codeforces rating increasing from 2046 to 2121, indicating enhanced problem-solving capabilities in competitive programming scenarios.
The efficiency gains are substantial, with 2-3x speed improvements for long-context inference tasks and 30-40% reductions in memory usage. These technical improvements translate directly into practical benefits for users, with API costs being reduced by more than 50% immediately upon release.
DeepSeek has made the core GPU kernels available as open-source under an MIT license, implemented in both CUDA and TileLang. This open approach facilitates rapid research prototyping and community contributions, accelerating innovation in sparse attention techniques across the AI ecosystem.
The model’s architecture retains the proven Multi-head Latent Attention (MLA) and Mixture-of-Experts (MoE) components from DeepSeek V3, with extensive pre-training on 14.8 trillion tokens. The sparse attention implementation enables efficient handling of long sequences while preserving the model’s reasoning capabilities across various domains including mathematics, coding, and general knowledge tasks.