DeepSeek’s V3 model was developed over two months for US$5. 58 million, utilizing fewer computing resources than its competitors. Reading Time: 2 minutes Why you can trust SCMP Reported by Ben Jiang in Beijing Published and Updated: 6:45pm, 27 Dec 2024 DeepSeek, a Chinese start-up, has stirred the global AI industry with its new large language model (LLM), which has outperformed models from Meta Platforms and ChatGPT creator OpenAI in benchmark tests. The Hangzhou-based company announced via WeChat on Thursday that its LLM, DeepSeek V3, boasts 671 billion parameters and was trained over roughly two months at a cost of US$5. 58 million, using notably fewer computing resources than those developed by larger tech firms. An LLM supports generative AI services like ChatGPT, and having a high parameter count is crucial for adapting to complex data patterns and making precise predictions. Computer scientist Andrej Karpathy, a founding team member at OpenAI, commented on the Chinese start-up's report on its new AI model, stating on social media platform X, “DeepSeek making it look easy …

with an open weights release of a frontier-grade LLM trained on a joke of a budget. ” Open weights imply releasing only the pretrained parameters, or weights, of an AI model, allowing third parties to use the model for inference and fine-tuning but not providing the training code, original data set, architecture details, and training methodology. DeepSeek’s creation of a strong LLM on a budget far smaller than what bigger companies like Meta and OpenAI usually invest highlights the progress made by Chinese AI firms, despite US sanctions restricting their access to advanced semiconductors necessary for training models.