中国大模型首登Nature封面！DeepSeek首次披露：R1训练只花了200万

Core Insights - DeepSeek has become the first Chinese large model company to be featured on the cover of Nature, with founder Liang Wenfeng as the corresponding author [2][3] - The R1 model has been recognized for its innovative approach, achieving significant performance improvements in reasoning tasks through a pure reinforcement learning framework [19][20] Group 1: Achievements and Recognition - DeepSeek's R1 model is the first large language model to undergo peer review, marking a significant milestone in the field [5] - The model has garnered 3,596 citations on Google Scholar and has been downloaded 10.9 million times from Hugging Face, indicating its widespread acceptance and use [7] - The training cost of R1 is approximately $294,000, significantly lower than competitors that often exceed $10 million, challenging the notion that high investment is necessary for top-tier AI models [12][13] Group 2: Training and Data - R1 was trained using 512 H800 GPUs for 198 hours, with a total training cost of $294,000 [10][11] - The dataset for R1 includes five types of data: Math, Code, STEM, Logic, and General, with a total of 126,000 prompts [15][18] - The model's training involved a combination of cold-start data, reinforcement learning, and supervised fine-tuning, enhancing its reasoning capabilities [25][26] Group 3: Performance Metrics - DeepSeek-R1-Zero achieved a pass@1 score of 71.0% in AIME 2024, significantly improving from 15.6% [21] - In comparison to other leading models, DeepSeek-R1 demonstrated competitive performance across various benchmarks, including MATH-500 and LiveCode [23][30] - The distilled models from DeepSeek-R1 outperformed direct applications of reinforcement learning on the base model, showcasing the effectiveness of the training approach [29] Group 4: Safety and Transparency - DeepSeek has released a detailed safety assessment of the R1 model, indicating a moderate inherent safety level comparable to GPT-4o [18][22] - The company has embraced transparency by open-sourcing the model weights for DeepSeek-R1 and DeepSeek-R1-Zero on Hugging Face, promoting community engagement [30]