China’s DeepSeek unveils new AI model that could halve usage cost
DeepSeek called the model the an advancement in its next-generation lineup of AI.
Chinese AI developer DeepSeek has released an experimental large language model that it says has much better training and reasoning, and which can be operated at a lower cost.
The Hangzhou-based company said its latest offering uses a “sparse attention" technique that cuts application programming interface prices by half. The API is the online interface that lets developers and companies access AI models and pay per use.
DeepSeek called the model the an advancement in its next-generation lineup of AI, in an article on Hugging Face, a developer forum, late Monday.
Chinese tech firms have been stepping up efforts to upgrade their LLMs as competition rises both at home and abroad. Last week, Alibaba Group rolled out a version of its flagship AI model that it described as its largest and most capable yet.
Western tech giants such as Google and OpenAI have also explored the “sparse attention" technique, which allows AI models to handle large amounts of information more efficiently. OpenAI wrote in 2019 that computing a full-attention matrix can be impractical for very large inputs, so “sparse patterns" where each output only considers a subset of inputs, is a better technique.
In a research paper released with the new model, DeepSeek said that its new model uses a “lightning indexer" and a “fine-grained token selection mechanism" to ensure attention is only applied to selected tokens.
Huawei Cloud said in a post late Monday that it has “quickly completed the adaptation" of the new DeepSeek-V3.2-Exp model.
DeepSeek’s V3.1 model and Alibaba’s Qwen3 series are the top-rated Chinese entries in the Artificial Analysis rankings of LLMs, behind offerings from players including OpenAI, xAI and Anthropic.
Write to Tracy Qu at tracy.qu@wsj.com
topics
