siliconflow/deepseek-v3.2

DeepSeek-V3.2 is a high-computational-efficiency model with superior reasoning and agent capabilities. It integrates three key breakthroughs: DeepSeek Sparse Attention (DSA), an efficient attention mechanism significantly reducing computational complexity while maintaining performance, optimized for long-context scenarios; a scalable reinforcement learning framework achieving performance comparable to GPT-5, with high-compute versions rivaling Gemini-3.0-Pro in reasoning; and a large-scale agent task synthesis pipeline integrating reasoning capabilities into tool usage scenarios for enhanced instruction-following and generalization in complex interactions. The model achieved gold medal performance at the 2025 IMO and IOI.