Deepseep got an update. The Chinese announce a new generation

The Chinese startup Deepseek updated the experimental model of artificial intelligence, which he described as a step towards new generation artificial intelligence – Bloomberga agency said on Tuesday.

The Chinese startup presented Deepseek-V3.2-EXP, explaining that it uses a new technique, which is called Deepseek Sparse Attention (DSA). The Chinese company said that it may reduce calculation costs and increase the efficiency of some models. The startup with Hangzhou said in a message that the latest version means an intermediate step towards the new generation structure, indicating that it works on this model with Chinese chip manufacturers.

The latest version is based on the older V3.1-Terminus, introducing a mechanism designed for exploration and optimization of training, as well as artificial intelligence. Startup announced that it aims to present his research on how to improve the performance of long text sequences.

The use of the Sparse Attention Mechanism by Deeps indicates Deepseek search for ways to train AI models, taking into account the limited access to NVIDIA Corp integrated circuits. and others. The founder of Deepseek, Liang Wenfeng, was a co -author of the article on this subject this year, in which it was described how developers can combine programming innovations with calibrated equipment to reduce the demand for computing power.

“Simply put, this means that the company devotes some accuracy of results, but tries to maintain a high level of intelligence. Continuous innovations in the performance of models would accelerate AI adoption and provide a better return on investment in China, despite restrictions on integrated circuits,” wrote Jefferies, Edison Lee.

According to Bloomberg, cost reduction gives Deepseek more freedom of competing. The company announced a reduction of its program tools by half the prices, By joining other Chinese startups that cut costs to attract users. On Monday Huawei Technologies Co. and Cambricon Technologies Corp. – Leaders of the Chinese market of AI systems – announced that their products will support the latest update of the Deepseek.

Deepseek has announced that the latest versions of its models support the FP8 or Floating Point 8 structure, while working on BF16 support. Both technical terms mean ways to store numbers on computers in the context of artificial intelligence and machine learning. In theory, FP8 saves memory and accelerates calculations.

AI models process millions of numbers, and using smaller formats such as FP8 and BF16, combines speed with accuracy and facilitates the launch of large models on less advanced equipment. Although the FP8 is not very precise, it is considered useful in many AI tasks. BF16 or Brain Floating Point 16 is considered more accurate when training AI models.

Chinese shares of companies listed in continental China dealing with semiconductors increased by 2 percent. After Deepseek released a new model. (PAP Biznes)

KEK/ ANA/