DeepSeek Unveils V3.2 Models With Record-Breaking Math Reasoning
DeepSeek releases two new AI models that offer advanced reasoning and cost efficiency, alongside a math model that outperforms top human competitors.
DeepSeek launches two new AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, to deliver significant upgrades in reasoning and autonomous agent performance. The standard V3.2 model provides GPT-5 level performance across web, mobile, and API platforms while maintaining impressive cost-efficiency. Meanwhile, the Speciale variant rivals Google's Gemini-3.0-Pro in raw capability, though it currently remains accessible only via API for a limited community evaluation period.
The most remarkable breakthrough comes from the separate DeepSeekMath-V2 model, which achieves an extraordinary 118 out of 120 points on the William Lowell Putnam Mathematical Competition. This score easily surpasses the top human score of 90 and demonstrates genuine mathematical reasoning rather than simple pattern matching. The system also performs at gold-medal levels in the International Mathematical Olympiad 2025.
This mathematical prowess stems from a revolutionary self-verifiable reasoning system that incorporates a proof generator, a dedicated verifier, and a meta-verification layer. These components work together in a continuous feedback loop to progressively improve the model's logical capabilities. Together, these releases show that DeepSeek pushes the boundaries of AI reasoning while keeping compute costs remarkably low.