DeepSeek Launches Open-Source Reasoning Models: DeepSeek-V3.2 and V3.2-Speciale
On December 1, Hangzhou’s DeepSeek unveiled two open-source reasoning models—DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. Marketed as delivering GPT-5-level performance, these models demonstrate exceptional capabilities in math and informatics contests. V3.2 focuses on efficiency for daily use, while the high-compute Speciale variant excels in reasoning tasks.
Founded in 2023 by Liang Wenfeng and funded by High-Flyer, DeepSeek prioritizes performance and cost-effectiveness. The models achieved impressive benchmark results, with V3.2-Speciale scoring 96.0% on the AIME 2025, outperforming both GPT-5 High and Gemini 3 Pro.
Technical innovations include the DeepSeek Sparse Attention (DSA) mechanism, enhancing long-context processing efficiency. This release aligns with the growing trend of Chinese open-source models gaining global traction, driven by rapid development cycles and U.S. export regulations.
DeepSeek has made V3.2 and its Speciale API available on Hugging Face, allowing for continued exploration and community contributions.
