Overview

• DeepSeek open-sourced the experimental V3.2-Exp model on Hugging Face and Alibaba-backed ModelScope, with downloads also available on its site and app.
• The company says V3.2-Exp improves training and inference efficiency and cuts API costs by more than 50% versus prior versions.
• DeepSeek outlined a DeepSeek Sparse Attention (DSA) approach in its V3.1-Exp research as an intermediate step toward a next-generation architecture.
• Huawei said its products will support DeepSeek’s latest update, and the new versions include FP8 support with work underway on BF16.
• The release follows a rapid cadence that saw V3.1-Terminus arrive a week earlier, and permissive licensing has driven extensive third-party forking on platforms like Hugging Face.