DeepSeek-V3.2-Exp

What it is

DeepSeek-V3.2-Exp is an experimental model that DeepSeek released on 29 September 2025. It was built on DeepSeek-V3.1-Terminus, and the "Exp" in its name marks it as an experiment rather than the final model of a new generation^[1]. Its purpose was to put a new attention mechanism into a working model and check what it cost and what it gave up, while keeping the rest of the model close to the version it came from^[1].

Sparse attention

The main change is DeepSeek Sparse Attention (DSA), which DeepSeek describes as a fine-grained sparse attention mechanism^[1].

Attention is the operation by which a language model lets each token in a piece of text draw information from other tokens. In the standard, dense form, every token is compared with every other token it is allowed to see, so the amount of computation and memory needed grows quickly as the context gets longer. Sparse attention keeps the same idea but lets each token look at only a selected subset of the other tokens, so the cost grows more slowly with length. "Fine-grained" refers to how that subset is chosen: the selection is made at a finer granularity than simply discarding whole chunks of the context.

DeepSeek's stated aim with DSA was to make long-context training and inference faster and cheaper while affecting output quality as little as possible^[1]. On quality, the company reported performance on par with V3.1-Terminus, the model it was built on^[1].

Release and pricing

The model went live in DeepSeek's app, on the web and through the API on the same day it was announced^[1]. Alongside it, API prices were cut by more than 50%, effective immediately^[1]. Because the new attention mechanism was presented as a way to lower the cost of long-context work, the price change and the architecture change arrived together as one story about serving cost.

For anyone who wanted to check the claim rather than take it, V3.1-Terminus stayed available through a temporary API until 15 October 2025, so the two models could be compared side by side^[1].

Open weights and kernels

The weights were open-sourced on Hugging Face, together with a technical report on GitHub^[1]. DeepSeek also released the key GPU kernels in both TileLang and CUDA, and recommended TileLang for rapid research prototyping^[1]. Releasing the kernels matters for a mechanism like sparse attention: the savings only appear if the low-level code that runs on the GPU actually exploits the sparsity, so the kernels are part of the claim, not an afterthought.

Where it sits

The experimental label turned out to be a staging step rather than a dead end. Its base model, DeepSeek-V3.2-Exp-Base, was later used as the starting point for DeepSeekMath-V2^[3]. And DeepSeek-V3.2 and V3.2-Speciale, released on 1 December 2025, kept the same architecture as V3.2-Exp^[4]. In other words, the sparse attention design first shipped in an experiment became the architecture of the models that followed it.