DeepSek Releases "Intermediate" AI Model Towards Next Generation Architecture

DeepSek, has released its newest "experimental" model. (photo; x @elandroidelibre)

JAKARTA An artificial intelligence (AI) developer from China, DeepSee, has released its latest "experimental" model. The model is claimed to be more efficient in training and better at processing long-text sequences than previous models.

The Hangzhou-based startup named the model DeepSek-V3.2-Exp, and called it a "step between towards a next-generation architecture" in a post on the Hugging Face developer forum.

The new architecture in question is likely to be the launch of DeepSek's most important product since the V3 and R1 models that surprised Silicon Valley and technology investors outside China.

The V3.2-Exp model is equipped with a mechanism called DeepSekSparse Attention, which the company says can cut computing costs while improving model performance in several aspects. In a post on platform X on Monday, September 29, DeepSek also announced that they cut API prices by more than 50%.

Although the next generation architecture of DeepSek is not expected to shock the market like the previous version in January, its success can still put great pressure on domestic competitors such as Alibaba's Qwen, as well as international players like OpenAI, if DeepSek is again able to display high performance at a much lower cost than its competitors.