DeepSeek V3.2 introduces Native Sparse Attention architecture designed to make long-context models more efficient without sacrificing performance. Instead of applying sparsity only during inference (after training is complete), NSA is designed to be sparse from the very beginning and is trainable from end-to-end.

By learning the sparse patterns during pretraining, DeepSeek is able to exceed the performance of standard Full Attention models in different benchmarks. It also allows the model to be efficiently fine-tuned for complex tasks like chain-of-thought reasoning.

There's a dramatic reduction in the number of token needed for both prefilling and decoding as the context length increases, making it much more economical to run.

• https://arxiv.org/pdf/2502.11089
• https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf

DeepSeek V3.2 introduces Native Sparse Attention architecture designed to make long-context models more efficient without sacrificing performance. Instead of applying sparsity only during inference (after training is complete), NSA is designed to be sparse from the very beginning and is trainable from end-to-end.

By learning the sparse patterns during pretraining, DeepSeek is able to exceed the performance of standard Full Attention models in different benchmarks. It also allows the model to be efficiently fine-tuned for complex tasks like chain-of-thought reasoning.

There's a dramatic reduction in the number of token needed for both prefilling and decoding as the context length increases, making it much more economical to run.

• https://arxiv.org/pdf/2502.11089
• https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf

DeepSeek V3.2 introduces Native Sparse Attention architecture designed to make long-context models more efficient without sacrificing performance. Instead of applying sparsity only during inference (after training is complete), NSA is designed to be sparse from the very beginning and is trainable from end-to-end.

By learning the sparse patterns during pretraining, DeepSeek is able to exceed the performance of standard Full Attention models in different benchmarks. It also allows the model to be efficiently fine-tuned for complex tasks like chain-of-thought reasoning.

There's a dramatic reduction in the number of token needed for both prefilling and decoding as the context length increases, making it much more economical to run.

• https://arxiv.org/pdf/2502.11089
• https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf