What is DeepSeek-V3.2, and how does sparse attention make long inputs cheaper?

Yesterday, Monday, December 1, 2025, the Chinese AI lab DeepSeek released two models: DeepSeek-V3.2, now the default in its app, website and API, and DeepSeek-V3.2-Speciale, a variant tuned for maximum reasoning that is available only through the API for now. Both are open-weight: the trained model files are published on Hugging Face, so anyone with enough hardware can download and run them.

V3.2 is the finished version of V3.2-Exp, an experimental model DeepSeek released in late September to test a new way of reading long text. With this release, that technique moves from experiment to the main product.

What DeepSeek released

DeepSeek describes the two models this way:

Speciale scored at gold-medal level in two 2025 contests for students, the IMO for mathematics and the IOI for programming, and DeepSeek also cites results at China's national math olympiad and the ICPC World Finals programming contest. The Decoder reports it placed tenth at the IOI and second at the ICPC World Finals.

V3.2 is also DeepSeek's first model that can think while using tools. A tool here is something the model can call, such as a web search or a code runner. In practice, V3.2 can reason, call a tool, read the result and keep reasoning within one answer, and it can also use tools in its faster non-thinking mode.

How sparse attention works

To understand the main change, start with attention, the core operation of a transformer, the design behind nearly all modern language models. Text is split into tokens, small pieces of words. When the model produces each new token, attention lets it look back at all the earlier tokens and decide which ones matter.

That full look-back is expensive. If the text is twice as long, each token has twice as much to look at, and there are twice as many tokens, so the work roughly quadruples. Researchers write this as O(L²), where L is the length. For a 128,000-token document, the cost adds up quickly.

DeepSeek Sparse Attention (DSA) adds a shortcut in two steps:

1. A small, fast component called the lightning indexer gives every earlier token a quick relevance score for the current one.
2. The model then pays full attention only to the top-scoring tokens, 2,048 of them in DeepSeek's training setup, and ignores the rest.

An analogy: instead of rereading an entire book every time you write the next sentence of a summary, you check the index, open the few pages that matter and read those carefully. The indexer still glances at everything, but its glance is very cheap. DeepSeek's report says this cuts the main attention cost from growing with the square of the length to growing with the length times that fixed number of selected tokens, giving large speedups on long inputs, while tests showed no real loss of quality compared with the previous version.

Training for reasoning and agents

The technical report names two more ingredients. First, DeepSeek spent more than 10% of its pre-training compute on post-training, the reinforcement learning stage after the model first learns from text. The Decoder notes that this share was around one percent two years ago. Second, to teach the model to act as an agent, an AI that pursues a goal through several steps and tools, DeepSeek generated training tasks in more than 1,800 environments with over 85,000 complex instructions.

Here is the kind of task that uses both long input and tool use:

Here are 30 customer support transcripts from last month. Find the
five most common problems, quote one short example of each, and use
the code tool to count how many transcripts mention each problem.
Show the results in a table.

Scores and limits

DeepSeek's own figures, as reported by The Decoder, put V3.2 close behind GPT-5 (High) in math and coding contests: 93.1% against 94.6% on AIME 2025, a competition math exam. On SWE Multilingual, a test of fixing real bugs in several programming languages, V3.2 reaches 70.2% against GPT-5's 55.3%. Gemini 3 Pro still leads most categories.

DeepSeek is open about the gaps. Its report lists three: the model's world knowledge is narrower than that of the best proprietary models, because it was trained with less total compute; it is less token-efficient, often needing longer answers to match their quality; and it still trails them on the hardest tasks. Speciale shows the trade-off clearly: The Decoder reports it used about 77,000 tokens on average for a programming contest problem, against roughly 22,000 for Gemini 3 Pro, which is why the everyday V3.2 has stricter length limits.

Why it matters

DeepSeek's report argues that in recent months closed models, available only through their makers' services, have been pulling further ahead of open ones. V3.2 is an attempt to close it while keeping costs low, and sparse attention makes long documents and long agent sessions cheaper to run. Releasing the weights also means other companies and researchers can study and adapt it.

How can you try DeepSeek-V3.2 and V3.2-Speciale?

The easiest way is DeepSeek's own app or website, where V3.2 is now the default model. Developers can use it through DeepSeek's API in the same way as V3.2-Exp; the API now supports thinking during tool use.

Speciale is only available through a temporary API endpoint until December 15, 2025, at the same price as V3.2 and without tool calls. It suits hard math or programming problems where you can afford long answers.

If you have powerful hardware, you can download both from Hugging Face and run them yourself. As with any model, check important answers, and remember that data typed into a hosted service is processed on that company's servers; read its terms before sharing anything sensitive.