Chinese AI developer DeepSeek has launched its latest experimental model, DeepSeek-V3.2-Exp, on September 29, 2025, marking a significant advancement in large language model technology. Built on the V3.1-Terminus foundation, this new release introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that dramatically boosts efficiency for long-context processing while maintaining output quality comparable to prior top models.
DSA technology allows the model to selectively focus on the most relevant portions of lengthy text sequences, reducing computation costs by 30-40% and improving training speed by around 50%. This innovation enables handling extremely long inputs up to 128,000 tokens, cutting memory use and inference time substantially.
Accompanying the technical breakthrough, DeepSeek has also slashed its API pricing by over 50%, making advanced AI capabilities more accessible with input costs as low as $0.07 per million tokens for cache hits. This aggressive price cut follows the company’s January 2025 market disruption with its R1 model and signals increased competitive pressure on both domestic rivals like Alibaba’s Qwen and global leaders such as OpenAI.
DeepSeek-V3.2-Exp is now available via the firm’s app, web platform, and APIs. The company has open-sourced the model, technical documentation, and GPU kernels to the AI research community through Hugging Face and GitHub, encouraging broad adoption and further innovation.
The release of V3.2-Exp underlines DeepSeek’s strategic push in the AI race, positioning it as a formidable contender in the global landscape by focusing on cost-efficiency and scalability of next-generation language models.
