Beijing (Reuters) – Chinese artificial intelligence developer has launched her latest “experimental” model, which, she said, is more efficient for training and better in processing long text sequences than previous versions.
The Hangzhou-based company called Deepseek-V3.2-EXP “intermediate stage towards our next generation architecture” in a publication on the Hugging Face forum. This architecture will probably be Deepseek’s most important product launch since V3 and R1 versions have shocked Silicon Valley and technology investors.
The V3.2-Exp model includes a mechanism called Deepseek Sparse Attention, which, according to the Chinese company, can reduce computing costs and increase the performance of some types of models. Deepseek said on Monday on social network X that it is cutting the API prices “over 50%”.
Free tool
XP Simulator
Learn in 1 minute how much your money can yield
Although it is unlikely that DeepSek’s next-generation architecture shakes markets as previous versions have done in January, it can still exert significant pressure on national rivals such as Alibaba’s QWEN and US counterparts such as OpenAi if you can repeat the success of Deepseek R1 and V3.
This will require the company to show high capacity for a fraction of what competitors charge and spend on model training.
(By Eduardo Baptista)
