deepseek-ai/DeepSeek-V3.2-Exp

DeepSeek-V3.2-Exp is an experimental model built on top of V3.1-Terminus, marking the first DeepSeek release to use DeepSeek Sparse Attention (DSA). DSA applies fine-grained sparse attention that the developers describe as having minimal impact on output quality while accelerating both training and inference on long sequences, with the practical benefit of reducing compute cost for long-context workloads.

DeepSeek reports that V3.2-Exp matches V3.1-Terminus on its published benchmarks, suggesting the new attention mechanism preserves quality while improving efficiency. Alongside the release, DeepSeek cut API prices by more than half and kept V3.1-Terminus available through a temporary comparison endpoint for side-by-side evaluation. The combination of sparse attention, benchmark parity with its predecessor, and substantially lower API pricing positions V3.2-Exp as a cost-effective option for long-context reasoning, code assistance, and tool-driven applications.

SiliconFlowdeepseek-ai/DeepSeek-V3.2-Expdeepseek

Open provider page Provider docs

Quick Info

Provider: SiliconFlow
Model key: deepseek-ai/DeepSeek-V3.2-Exp
Release date: Oct 10, 2025
Last updated: Nov 25, 2025
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.27
Output token cost: $0.41

Limits

Output tokens: 164,000 tokens
Context window: 164,000 tokens

Transparent token rates

Compare deepseek-ai/DeepSeek-V3.2-Exp pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about deepseek-ai/DeepSeek-V3.2-Exp

SiliconFlow

Coverage

DeepSeek's first-party API documentation announces DeepSeek-V3.2-Exp as an experimental model built on V3.1-Terminus, debuting DeepSeek Sparse Attention (DSA) for fine-grained sparse attention that improves long-context training and inference efficiency. According to the page, benchmarks show V3.2-Exp performing on par The announcement also lists open-source release artifacts including the Hugging Face model, the technical report PDF on GitHub, and key GPU kernels implemented in TileLang and CUDA. For comparison testing, DeepSeek states that V3.1-Terminus remains accessible via a temporary API endpoint until October 15, 2025 at 15:59

SiliconFlow

Official sourceBenchmark

Compare DeepSeek-V3.2 and DeepSeek-V3.2-Exp across performance, cost, capabilities, and real-world use cases. See which model fits your needs.

SiliconFlow

CoverageBenchmark

A deep technical breakdown of DeepSeek V3.2, examining how training data, synthetic pipelines, sparse attention, and post-training RL shape reasoning and performance.