Chinese AI developer DeepSeek has unveiled an “experimental” model that it says is more efficient to train and better at handling long sequences of text than its previous large language models.
The Hangzhou-based company described the new model, DeepSeek-V3.2-Exp, as an “intermediate step toward our next-generation architecture” in a post on the developer platform Hugging Face. That upcoming architecture is expected to be DeepSeek’s most significant release since its V3 and R1 models shook Silicon Valley and global tech investors earlier this year.
A key feature of V3.2-Exp is DeepSeek Sparse Attention, a mechanism designed to cut computing costs while improving certain types of performance. The company also announced via X on Monday that it is reducing its API prices by more than 50%.
Although the new model is unlikely to trigger the same market reaction as DeepSeek’s January launches, analysts say it could still put pressure on rivals in China, such as Alibaba’s Qwen, as well as U.S. competitors like OpenAI, if it delivers high performance at significantly lower training costs.
