DeepSeek postpones launch of new AI

DeepSeek has had to delay the release of its R2 model due to performance that has not met expectations, while China's AI chip supply continues to be affected by US restrictions.

DeepSeek postpones launch of R2 model. Photo: Bloomberg .

DeepSeek, a Chinese artificial intelligence startup, has yet to officially release its next-generation language model, R2. According to a report from The Information on June 26, the main reason is that CEO Liang Wenfeng is not satisfied with the current performance of the model the team is developing.

R2 is expected to be a significant upgrade to the R1 model, which has made a splash globally thanks to its superior inference and programming capabilities. Reuters previously reported that DeepSeek plans to launch R2 in May. The goal is to improve its ability to handle diverse languages ​​amid growing demand for AI localization.

Over the past several months, DeepSeek’s engineering team has been steadily refining the R2 model. But the decision on when to release it remains CEO Liang’s. Insiders say he is being cautious about every aspect of the product, from technical performance to commercial viability, to avoid unnecessary risk.

In addition to internal factors, DeepSeek is also facing external challenges. Specifically, the shortage of AI chip supply makes it difficult for the company to operate due to limited hardware power. Chinese cloud computing companies, which are partners in distributing models to corporate customers, are struggling due to recent US technology export restrictions.

According to The Information , most of its current customers are running R1 models on Nvidia’s H20 chip platform. However, since April, the chip line has been placed on the US government’s export restriction list, which is also the only AI chip line that Nvidia can legally export to China at that time.

Meanwhile, DeepSeek appears to have contacted several major domestic cloud service providers and submitted technical specifications to prepare the infrastructure for R2. However, the scalability of the model remains unclear, as it depends on both the leadership’s decisions and the global AI chip market situation.

Source: https://znews.vn/deepseek-hoan-ra-mat-ai-moi-post1564148.html