Chinese AI developer DeepSeek has unveiled an experimental model, DeepSeek-V3.2-Exp, promised to offer enhanced training efficiency and superior processing of lengthy text sequences.
In a recent post on the developer forum Hugging Face, the Hangzhou-based company revealed that this model represents an “intermediate step toward our next-generation architecture.” This architecture is hailed as DeepSeek’s most crucial launch since previous versions like V3 and R1 made headlines in Silicon Valley and appealed to tech investors globally.
The V3.2-Exp model features DeepSeek Sparse Attention, a mechanism designed to slash computing costs while enhancing certain model performances. DeepSeek announced on platform X that it would significantly reduce API prices by over 50%. While the upcoming architecture may not disrupt markets as its predecessors did, it still holds the potential to apply competitive pressure on domestic adversaries such as Alibaba’s Qwen and international rivals like OpenAI, provided it can achieve comparable breakthroughs as DeepSeek R1 and V3.
(With inputs from agencies.)