Official announcements from DeepSeek or the partnering US tech company

DeepSeek Launches DeepSeek-V3, Surpassing GPT-4o with 671B Parameters, 37B Activated, Trained on 14.8 Trillion Tokens for $5.5 Million

Dec 27, 2024, 02:31 AM

Chinese AI company DeepSeek has launched DeepSeek-V3, a new open-source language model that reportedly surpasses OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet in performance. The model features a total of 671 billion parameters, with 37 billion activated for each token, and was trained using 14.8 trillion tokens at a cost of approximately $5.5 million. This achievement is notable given that it was developed with a fraction of the resources typically required for such advanced models. DeepSeek claims that its model is more efficient and cost-effective, having been trained on 2,048 GPUs. The company, which has not sought outside funding, operates entirely on self-generated capital from a hedge fund. The release of DeepSeek-V3 is seen as a pivotal moment in the AI landscape, particularly in the context of ongoing geopolitical tensions and discussions around AI regulation in the United States.