DeepSeek debuts sparse attention in DeepSeek-V3.2-Exp and halves API prices

Background

DeepSeek-V3.2-Exp is an experimental model from DeepSeek, a Chinese AI company. It was built on DeepSeek-V3.1-Terminus, an earlier model in the same series, and was released on 29 September 2025 ^[1]. Its purpose was to test a new way of handling long inputs rather than to launch a new model family.

Long inputs are expensive for transformer models. In standard attention, each token in a text is compared with the tokens before it, so the work grows quickly as the context gets longer. Context is the amount of text a model can consider at one time. People who work with long documents, large codebases or long multi-step conversations pay that cost during both training and inference.

What happened

DeepSeek released V3.2-Exp with a new mechanism called DeepSeek Sparse Attention, abbreviated DSA ^[1]. Sparse attention is a family of techniques in which a model does not compare every token with every other token. It selects a subset of earlier tokens to attend to. DeepSeek described DSA as a fine-grained sparse attention mechanism, meaning the selection works at a detailed level rather than in large blocks. The stated goal was to make long-context training and inference faster and cheaper, with minimal effect on the quality of the output ^[1].

According to DeepSeek, the model performed on par with DeepSeek-V3.1-Terminus ^[1]. The release was presented as an experiment that tests the new mechanism, not as a finished flagship product.

Key details

The release brought three practical changes. First, DeepSeek cut the prices of its API by more than 50 percent, effective immediately ^[1]. The API is the interface that lets outside programs send requests to the model. Second, the weights, which are the learned numerical parameters that define how the model behaves, were published on Hugging Face, a platform for sharing machine learning models ^[1]. Third, a technical report was published on GitHub ^[1].

DeepSeek also released the key GPU kernels, the low-level programs that run calculations on graphics processors, in two versions: one written in TileLang and one in CUDA. CUDA is the programming platform used with Nvidia GPUs. DeepSeek recommended TileLang for rapid research prototyping, meaning quick experiments in which a researcher tries out new ideas ^[1].

Impact

V3.2-Exp is best understood as a test of DSA rather than as a general-purpose new model. Because the weights, the technical report and the kernels were all made public, other researchers could examine the approach and reuse parts of it. The price cut meant that developers using the API could run long requests at a lower cost from the first day ^[1].