DeepSeek-V3.2-Exp

Claims and evidence

• Released 29 September 2025 ● Primary source[1]page header 2025/09/29 [4]Date: 2025-09-29

• Status AvailableWeights are still public on DeepSeek's Hugging Face organisation under the MIT License. In the DeepSeek API it was replaced by DeepSeek-V3.2 on 2025-12-01. ● Primary source[2]License [4]Date: 2025-12-01
• Replaced by DeepSeek-V3.2In the DeepSeek API only: both API model names were upgraded to DeepSeek-V3.2, which DeepSeek calls the official successor to V3.2-Exp. The weights stay available. ● Primary source[4]Date: 2025-12-01 [5]

• Derived from (continued pretraining) DeepSeek-V3.1-TerminusContinued pre-training from a DeepSeek-V3.1-Terminus base checkpoint (context extended to 128K) to add DeepSeek Sparse Attention, followed by post-training. ● Primary source[3]Abstract; 1. Architecture; 2. Training [1] [2]Introduction

• Change · Architecture Adds DeepSeek Sparse Attention, a fine-grained sparse attention mechanism with a lightning indexer; DeepSeek says this is the only architectural change.Compared with DeepSeek-V3.1-Terminus ● Primary source[3]1. Architecture
• Change · Efficiency DeepSeek says sparse attention makes long-context training and inference more efficient; it cut API prices by more than 50 percent with this release.Compared with DeepSeek-V3.1-Terminus ● Primary source[1]

• Feature Reasoning modeThinking and non-thinking modes, served as deepseek-reasoner and deepseek-chat. ● Primary source[4]Date: 2025-09-29
• Open weights YesMIT License for the repository and the weights. ● Primary source[1]Open Source Release [2]License
• Access API, consumer app, Open-weights download ● Primary source[1]

Lineage

Variants

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from DeepSeek on AI Radar →