DeepSeek-R1发布，性能对标OpenAI o1正式版深度求索DeepSeek2025-0l-2020:l0:59今天，我们正式发布DeepSeek-R1,并同步开源模型权重。·DeepSeek-RI遵循MIT License,允许用户通过蒸馏技术借助R1训练其他模型。·DeepSeek-RI上线API,对用户开放思维链输出，通过设置'model-='deepseek-reasoner''即可调用.●DeepSeek官网与App即日起同步更新上线。性能对齐OpenAI-ol正式版DeepSeek-RI在后训练阶段大规模使用了强化学习技术，在仅有极少标注数据的情况下，极大提升了模型推理能力。在数学、代码、自然语言推理等任务上，性能比肩OpenAI ol正式版。DeepSeek-R1OpenAI-01-1217 DeepSeek-R1-32BOpenAI-o1-miniDeepSeek-V31005.280-多62.11020AIME 2024CodeforcesGPQA DiamondMATH-500MMLUSWE-bench Verified在此，我们将Deep Seek-RI训练技术全部公开，以期促进技术社区的充分交流与创新协作。论文链接：https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf蒸馏小模型超越OpenAI o1-mini我们在开源DeepSeek-RI-Zero和DeepSeek-.RI两个66OB模型的同时，通过DeepSeek-R1的输出，蒸馏了6个小模型开源给社区，其中32B和70B模型在多项能力上实现了对标OpenAI ol-mini的效果。AIMEAIMEMATH-GPQA20242024500CodeForcesDiamondGPT-4o-05139.313474.632.9759.0Claude-3.5-Sounet-102216.026.778.365.038.9717.0o1-mini63.680.090.060.053.81820.0QwQ.32B44.060.090.654.51316.0DecpScck-R1-Distill-Qwen-1.5B28.952.783.933.816.9954.0DeepSeek-R1-Distill.Qwen-7B55.592.849,137.61189.0DeepSeek-R1-Distill-Qwen-14B69.780.093.953.11481.0DeepSeek-R1-Distill-Qwen-32B72.683.394.362.11691.0DeepScek-R1-Distill-Llama-8B50.480.089.149.039.61205.0DeepSeek-R1-Distill-Llamna-70B70.086.794.565.257.51633.0HuggingFace链接：https://huggingface.co/deepseek-aio deepseek-ai/DeepSeek-R1-Distill-Llana-70Bdeepseek-ai/DeepSeek-R1-Distill-Owen-32Bodeepseek-ai/DeepSeek-R1-Distill-Qmen-148deepseek-ai/DeepSeek-R1-Distill-Qwen-7Bo dcepscck-ai/DcepScek-R1-Distill-Qwen-1.50deepseck-ai/DecpSerk-R1-Distil1-Llana-8Bdeepseek-a1/DeepSeek-Rldeepseek-a1/DeepSeek-R1-Zero开放的许可证和用户协议