Home / Technology / DeepSeek AI Challenges Giants with New Models
DeepSeek AI Challenges Giants with New Models
2 Dec
Summary
- DeepSeek released two new reasoning-capable AI models, V3.2 and V3.2-Speciale.
- The V3.2-Speciale model reportedly outperforms Gemini 3.0 Pro and GPT-5 High.
- DeepSeek commits to remaining open source, making its models accessible for free.
DeepSeek, an AI firm based in China, has unveiled two new reasoning-capable artificial intelligence models: V3.2 and V3.2-Speciale. The company announced that its V3.2-Speciale model has surpassed the performance of established models such as Gemini 3.0 Pro and GPT-5 High in benchmark tests conducted by DeepSeek. Currently, the V3.2 model is accessible through its app and web interface, while the more advanced V3.2-Speciale is available via API access.
DeepSeek V3.2 builds upon the experimental V3.2-Exp released previously. The company asserts that V3.2 matches the reasoning capabilities of Gemini 3.0 Pro and exceeds that of OpenAI's GPT-5 in reasoning benchmarks. Notably, DeepSeek has pledged to maintain an open-source approach, allowing any entity to utilize its models without charge. However, the significant parameter count of V3.2 suggests substantial hardware requirements for operation.
The introduction of DeepSeek's V3.2 models follows its earlier impact on the AI industry when it released a free, reasoning-capable model online. This move shifted the narrative around AI development, highlighting the potential for highly capable models at reduced costs. DeepSeek-V3.2-Speciale has also reportedly secured top honors in prestigious competitions like the International Mathematical Olympiad and International Olympiad in Informatics in 2025.
