DeepSeek has unveiled its new large language model, DeepSeek-V3.2-Exp, which requires less computing power and is described by the company as “experimental.” The company characterizes this release as “an intermediate step toward next-generation architectures.”
The main advantage of the new model is that it can be trained at a lower cost compared to previous versions. This will help DeepSeek reduce both development and usage costs of AI. Moreover, the model promises to be more capable in processing long texts, such as large documents or extended conversations.
The secret behind the model’s efficiency lies in DeepSeek’s Sparse Attention technology. This mechanism allows the model to focus only on the most important parts of the text instead of processing all the information. As a result, operational costs are significantly reduced while maintaining, and in some cases even improving, performance.
Right after this announcement, DeepSeek revealed that it had cut the prices of its API services by more than 50%. This is a major benefit for developers and companies seeking to use their technology.
Although the new model may not have as big an impact as DeepSeek’s earlier giants, R1 and V3, if successful, it could put serious competitive pressure on both Chinese rivals (such as Alibaba’s Qwen) and global AI leaders like OpenAI. This is because cheaper and more efficient models have the potential to reshape the rules of the AI industry.
