DeepSeek
recently announced the official release of two new production models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, marking another major milestone in the company’s pursuit of practical and high-performance AI reasoning systems. Both models are now part of DeepSeek’s core model lineup and represent the latest results of large-scale training, system optimization, and evaluation across real-world and academic benchmarks.
DeepSeek-V3.2 is designed as a general-purpose reasoning model that balances strong inference ability with efficient output length, making it suitable for daily professional and consumer use. In publicly available reasoning benchmarks, DeepSeek-V3.2 has demonstrated performance on par with GPT-5 and slightly below Gemini-3.0-Pro. Compared with Kimi-K2-Thinking, V3.2 significantly reduces output length while preserving reasoning quality, which lowers compute costs and shortens response time for users. This makes DeepSeek-V3.2 well-suited for production environments that require both performance and efficiency.
DeepSeek-V3.2-Speciale represents the company’s frontier research direction and focuses on pushing the upper limits of open-source model reasoning. As a long-thinking enhanced edition of V3.2, it integrates theorem-proving capabilities from DeepSeek-Math-V2 and is optimized for complex mathematical and algorithmic reasoning. The Speciale model has achieved top-tier results in multiple international competitions, including gold medals at IMO 2025, CMO 2025, ICPC World Finals 2025, and IOI 2025. These results demonstrate DeepSeek’s ability to combine foundational model training with domain-specific reasoning frameworks to reach performance previously limited to closed research systems.
DeepSeek-V3.2 is now available across the official web platform, mobile app, and API, enabling developers and users to immediately access the latest production model. DeepSeek-V3.2-Speciale is currently offered through a temporary API and is open to the community for evaluation and research purposes, supporting further exploration at the boundaries of reasoning intelligence.
With the launch of V3.2 and V3.2-Speciale, DeepSeek continues its commitment to delivering reliable, scalable, and transparent AI systems for both real-world application and frontier research.
If you like this article, please give us a like! 💖 👍