技术规格

系统要求

不同量化方法和上下文大小的显存要求

DeepSeek-V3.2 Thinking

DeepSeek-V3.2 Thinking is the reasoning-enhanced variant of DeepSeek-V3.2, specifically optimized for complex problem-solving through chain-of-thought reasoning. Based on the same 671B parameter MoE architecture with 37B activated parameters, this model is fine-tuned to produce detailed reasoning traces before generating final answers. Excels at multi-step logical reasoning, mathematical proofs, algorithmic problem-solving, and tasks requiring explicit step-by-step thinking. Achieves enhanced performance on reasoning benchmarks: 94.8% on MATH-500 (with reasoning), 85.2% on Codeforces, and 73.4% on AIME. The thinking mode provides transparency into the model's reasoning process, making it ideal for educational applications, research, debugging complex logic, and scenarios where interpretability is crucial. Supports 128k context window with strong multilingual reasoning capabilities. MIT licensed.