*** DeepSeek v3.2 Exp: Spatial Attention ***

DeepSeek-V3.2 Experimental - Key Specifications: 14 - 671B total parameters, 21B active per token 15 - 128K context length 16 - Sparse Mixture-of-Experts architecture 17 - Supports vision-to-language processing 18

**References:**
1. DeepSeek. (2024). DeepSeek-V3.2 Experimental Technical Report. DeepSeek AI.
https://github.com/deepseek-ai/DeepSeek-V3.2

2. DeepSeek-V3.2-Experimental. (2024). Model card and specifications. Hugging Face.
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Experimental

3. DeepSeek AI. (2024). "Sparse Mixture of Experts for Large Language Models"
Technical report detailing 671B parameter architecture with 21B active.