The Chinese developer of AI, Deepseek, launched its latest “experimental” model. According to the company, training is more efficient and long text sequences processing is better than the previous iterations of its large language models.
The company, based in Hangzhou, described Deepseek-V3.2-Exp as an “intermediate step towards our next generation architecture” in a publication in the Hugging Face developer forum.
It is likely that this architecture is the most important product launch of Deepseek since versions V3 and R1 surprised Silicon Valley and technological investors outside China.
You might interest you: the Chinese artificial intelligence model Depseek learns more when receiving ‘reward’
Depseek maintains pressure on Chinese rivals
The V3.2-Exp model includes a mechanism called Depseek Sprse Attention, which, according to the Chinese company, can reduce computer costs and improve the performance of some models. Deepseek announced Monday in an X post that is reducing API prices by more than 50 %.
While the next generation architecture of Deepseek is unlikely to revolutionize markets as the previous versions did in January, it could still exert significant pressure on national rivals such as Alibaba Qwen and its American counterparts as OpenAi if you can repeat the success of Deepseek R1 and V3.
That would require demonstrating a high capacity for a fraction of what its competitors charge and spend on models training.
With Reuters information
Do you like photos and news? Follow us on our Instagram
