DeepSeek has released the V3.2 and V3.2-Speciale models across web, app, and API. The company said V3.2 adds built-in reasoning for agent tasks and is its first model to support tool calls in both reasoning and non-reasoning modes. It reportedly reaches GPT-5-level results on public reasoning benchmarks while reducing output length and computational cost.
V3.2-Speciale is available through a temporary API for research and extends long-form reasoning and mathematical proof. The model achieved gold-level results in IMO, CMO, ICPC World Finals, and IOI 2025. DeepSeek said the Speciale version consumes more tokens and does not support tool use, while the regular V3.2 is optimized for daily tasks. [TechNode reporting]
Related
