At the end of September last year, DeepSeek quietly released V3.2 without much fanfare, but this version introduced a rather interesting architectural change.
The core change in V3.2 is not about parameter scale, but the attention mechanism. They built something called DeepSeek Sparse Attention (DSA), converting the traditional dense attention into a sparse version.
Where Sparse Attention Actually Saves
The traditional attention mechanism has O(L²) complexity — when the sequence length doubles, the computation quadruples. When your context window reaches 128K tokens, this overhead becomes very significant.
DSA works as follows:
- First, use the Lightning Indexer (FP8 precision) to quickly calculate the relevance scores between each query and historical tokens
- Then, select only the most relevant Top-k tokens for attention computation
- Complexity drops from O(L²) to O(Lk), where k is a fixed small number
Simply put, "not every word needs to look at all other words — just the most important ones are enough." This idea is not new, but DeepSeek made it work and put it into production.
Performance: Not Significantly Up, Not Significantly Down
To be honest: V3.2's benchmarks are roughly on par with V3.1. This is not a "stronger" release.
| Task | V3.2 Result |
|---|---|
| MMLU-Pro | 85.0 |
| AIME 2025 | 89.3 |
| Codeforces Rating | 2121 (up from 2046) |
| GPQA/HLE Reasoning | Slightly declined |
Coding ability did improve, with the Codeforces rating rising from 2046 to 2121 — a notable jump. However, reasoning benchmarks declined slightly because "generated reasoning tokens decreased" — one of the costs of sparse attention, as the generated chain-of-thought becomes shorter.
So V3.2 is a trade-off of efficiency for capability, not an across-the-board improvement.
Affordability Is the Selling Point
Pricing: $0.28/M input, $0.42/M output.
Cache hit: $0.028/M, 90% cheaper than cache miss.
Comparison:
| Model | Input (/M) | Output (/M) |
|---|---|---|
| GPT-5 | $1.25 | $10 |
| Claude Sonnet 4 | $3 | $15 |
| DeepSeek V3.2 | $0.28 | $0.42 |
Compared to GPT-5, the cost gap is nearly 5x. This is not the most powerful model, but if your scenario involves high-volume calls and cost-sensitive applications, V3.2's cost-effectiveness is hard to beat.
Context window: 128K tokens. Weights are on Hugging Face under the MIT license, free for commercial use.
Architecture Details
DSA is embedded in the Transformer base, running under MLA's MQA mode. Inference is primarily optimized for H800 GPU clusters; actual performance on other hardware may vary, and DeepSeek says broader validation is ongoing.
Training used unified GRPO reinforcement learning, combining reasoning alignment, instruction following, and agent tasks into a single training process.
Perspective
V3.2 is not a "smarter" model — it's a "cheaper" model.
The sparse attention mechanism is an engineering trade-off: sacrificing a bit of reasoning capability for substantial computational cost savings, then passing those savings directly to API prices.
For application developers, this approach is more practical than obsessing over benchmarks. Not every scenario needs the strongest model, but every startup cares about API costs.
DeepSeek's approach here differs from OpenAI and Anthropic — the latter tend to release the strongest models first and optimize costs later. DeepSeek brings costs down first, scales up user numbers, and then differentiates on top of that.
This playbook is common in China's tech industry, but seeing it work in the large model space is still worth paying attention to.
Sources: CocoLoop, DeepSeek V3.2-Exp Release: Pricing, API Costs, Context Window & Benchmarks (llm-stats.com); DeepSeek V3.2 Exp Model Specs, Costs & Benchmarks (Galaxy.ai); Top 10 Cheapest Providers for DeepSeek V3.2 in 2026 (DEV Community)