DeepSeek V3.2-Exp: The Sparse Attention Experiment, Explained

DeepSeek V3.2-Exp is easy to misread as a minor point release. It was not a capability upgrade at all, and DeepSeek said so plainly: the goal was to show that a new attention mechanism could make long-context inference much cheaper without making the model worse.

That framing matters if you are deciding whether to use it today. An experiment that succeeded gets folded into the next stable model, and that is exactly what happened here. Understanding what V3.2-Exp tested explains a good deal about why later DeepSeek models are priced the way they are.

Why an experimental release existed at all

DeepSeek released V3.2-Exp on 29 September 2025, describing it as an intermediate step towards its next-generation architecture. It was built on V3.1-Terminus, the model it sat next to, and went live on DeepSeek's app, web and API on the same day.

The weights went up on Hugging Face as a 685B-parameter checkpoint under the MIT licence, alongside a technical report and the GPU kernels needed to run the new attention path. DeepSeek also kept V3.1-Terminus available on a temporary API until 15 October 2025, specifically so people could run the two side by side.

That temporary endpoint is the tell. A lab that expects you to compare two models on your own workloads is telling you the change is meant to be invisible in output quality and visible only on the bill.

What sparse attention actually changes

Standard attention compares every token with every earlier token. Cost grows with the square of the sequence length, so a prompt that is four times longer costs roughly sixteen times as much attention compute. At 128K tokens that term dominates everything else. Attention mechanisms explained covers the mechanics if that is unfamiliar.

DeepSeek Sparse Attention (DSA) adds an indexer module that selects which earlier tokens are worth attending to for each new token, so the expensive attention step runs over a subset rather than the whole history. DeepSeek described it as fine-grained sparse attention, and the claim was that quality stays nearly identical while long-context cost falls substantially.

Two practical consequences follow:

• Long prompts get cheaper to serve. Prefill and decode both do less work per token once the context is large, which is where agent transcripts and repository-scale prompts live.
• Nothing changes for short prompts. If your typical request is a few thousand tokens, quadratic attention was never your main cost, and DSA buys you very little.

One implementation detail is worth knowing if you self-host. DeepSeek's model card later noted that the original inference demo code had a RoPE layout mismatch in the indexer, which could degrade output. It was fixed in a November 2025 update. If you benchmarked a self-hosted V3.2-Exp before then and found it disappointing, that may be why.

The benchmark table is the whole argument

For most releases a benchmark table is a sales pitch. For V3.2-Exp it is the experiment's result, and the interesting outcome is how little moved. DeepSeek's model card compares it with V3.1-Terminus:

• MMLU-Pro: 85.0 against 85.0
• GPQA-Diamond: 79.9 against 80.7
• Humanity's Last Exam: 19.8 against 21.7
• AIME 2025: 89.3 against 88.4
• Codeforces rating: 2121 against 2046
• SWE-bench Verified: 67.8 against 68.4
• Terminal-bench: 37.7 against 36.7

Some up, some down, nothing dramatic. That is what "on par" looks like, and it is the result DeepSeek wanted. The one row worth a second glance is Humanity's Last Exam, where the drop is larger than the others; if your workload is hard knowledge-heavy reasoning, test that specifically rather than assuming parity.

These are vendor-run numbers. Vendor-claimed vs independent benchmarks covers how much weight to give them, but for a parity claim the risk is smaller than usual: the vendor is not claiming to beat anyone, only to have not broken anything.

What it did to the price

The release came with a cut of more than 50% on DeepSeek's API prices, effective immediately. Launch coverage reported $0.28 per million input tokens on a cache miss, $0.028 on a cache hit and $0.42 per million output, against $0.56, $0.07 and $1.68 under V3.1-Terminus pricing.

The cut on output is the striking part, and it lines up with the efficiency story: decoding over a long context is where sparse attention saves the most. On OpenRouter the model is listed at around $0.27 input and $0.41 output per million tokens at the time of writing, with a 163,840-token context window. Prices drift, so check before you budget.

Should you still run it

For new work, almost certainly not. DeepSeek shipped the stable V3.2 on 1 December 2025 as the official successor, using the same model structure, and the V4 family has since moved the line much further. If you want the V3.2 generation, use DeepSeek V3.2; if you want cheap long-context work today, read the DeepSeek V4.1 Flash guide instead.

There are three narrow reasons to keep V3.2-Exp around:

1. Reproducibility. If you published results or built an evaluation baseline on it, keep the exact weights so you can rerun them.
2. Attention research. It is the cleanest public comparison of dense against sparse attention on an otherwise similar model, because V3.1-Terminus exists as the control.
3. A pinned production path you have not yet migrated. In that case, plan the move rather than discovering a hosted endpoint has quietly gone. Model deprecation and migration covers how to schedule it.

How to test a long-context efficiency claim yourself

The useful habit V3.2-Exp teaches is checking that an efficiency change really is quality-neutral on your own inputs. The method is short:

1. Take twenty real prompts and bucket them by length: under 8K, around 32K, and 100K or more.
2. Run each through the old and new model with identical settings and record output tokens, latency and cost.
3. Grade the outputs blind. If quality holds in the short bucket but slips in the long one, the sparse path is dropping something your task needs.
4. Compare cost per correct answer, not cost per token.

If the long bucket holds up and the bill falls, the efficiency change is real for you. If it does not, you have found the edge of the claim cheaply. How to benchmark LLMs on your own work has the fuller harness.