DeepSeek debuts new AI model as ‘intermediate step’ towards next generation

DeepSeek, a prominent player in the AI landscape, has introduced its latest model, DeepSeek-V3.2-Exp. This new iteration incorporates the innovative DeepSeek Sparse Attention technology, which the company describes as a tool for enhancing training and inference efficiency in scenarios involving long-context processing. The firm's announcement positions this model as a crucial milestone towards its next-generation architecture. Earlier this year, DeepSeek made waves in Silicon Valley by releasing its V3 and R1 models, which showcased performance levels comparable to industry giants like OpenAI and Google, all while significantly reducing development costs. Building upon the previous V3.1 model, DeepSeek-V3.2-Exp aims to optimize AI training and operations. The company emphasizes the model's ability to improve computational efficiency when dealing with lengthy text sequences. "This experimental release reflects our commitment to advancing transformer architectures, particularly in enhancing performance for extended contexts," DeepSeek stated on Hugging Face. In a recent update shared on X (formerly Twitter), DeepSeek highlighted that its DSA technology achieves "fine-grained sparse attention with minimal impact on output quality," ultimately enhancing long-context performance while decreasing computational demands. Furthermore, the company announced plans to cut API prices by 50% for global users. Although the new AI model isn't set for immediate global release, it stands to increase competition among Chinese AI firms, particularly with Alibaba's Qwen. Recently, Alibaba upgraded its Qwen model with the Qwen 3 Max, which reportedly surpassed OpenAI's GPT-5 in LMArena's benchmark tests.