According to people familiar with the matter, Chinese artificial intelligence company DeepSeek plans to release the latest generation of large language model V4 next week. This is the company’s first major update since the launch of its last blockbuster product more than a year ago. It is regarded as an important step for China to continue to challenge American competitors in the field of artificial intelligence.
According to two people familiar with the situation, DeepSeek is headquartered in Hangzhou, and the V4 launched this time will be a multi-modal model with image, video and text generation capabilities. Multiple people familiar with the matter said that DeepSeek has cooperated with China's local AI chip manufacturers Huawei and Cambrian to customize and optimize V4 to adapt to the latest generation chip products of both parties, thereby forming closer collaboration at the computing power level. This move is seen as another sign that Chinese technology companies are accelerating away from dependence on Nvidia's high-end AI chips, which are currently subject to U.S. export controls and related measures aimed at curbing China's technological rise.
The timing of this release is also quite symbolic. DeepSeek plans to launch V4 on the eve of China’s annual national “Two Sessions”, which this year will open on March 4. This high-profile political meeting provides the company with an important window of exposure and may further solidify its image as a "national AI champion."
This is the first major version iteration of DeepSeek since the release of the R1 inference model in January 2025. At that time, the company claimed that it had trained a system that was comparable in capability to leading models with a computing power scale that was far lower than that used by leading companies in Silicon Valley. This news once caused shock in the U.S. technology stock market. Some analysts described it as a "Sputnik moment" that marked China's rapid catching up and even rewriting of the landscape in the field of artificial intelligence. Since then, DeepSeek has launched more incremental updates than a complete new architecture, which has also given domestic competitors including Alibaba and Moonshot additional room for growth in the low-cost, open-source Chinese model market.
Multiple people familiar with the matter predict that DeepSeek's optimization of V4 specifically for domestic AI chips will help boost market demand for local chips and accelerate the transfer to Chinese manufacturers such as Huawei and Cambrian in the model inference stage (that is, the process of using trained models to generate answers), reducing dependence on Nvidia and AMD chips. Reuters had previously reported the progress of DeepSeek’s cooperation with Huawei and Cambrian. Another person familiar with the situation said that DeepSeek is not working with Nvidia on V4 optimization.
However, in the field of model training, NVIDIA still dominates, especially in the pre-training stage that requires huge amounts of computing power, and its GPU is still the industry standard. The Financial Times previously reported that DeepSeek tried to complete this initial training on Huawei hardware, but encountered technical difficulties in the process. When the company released the R1 model last year, it also published a detailed technical report explaining how to train and run the model more efficiently on NVIDIA chips. The relevant engineering methods have received widespread attention and praise. Some insiders believe that DeepSeek's sharing of its training methods for building "inference models" actually provides other laboratories with a reusable engineering path, helping the latter improve model inference capabilities under limited computing power.
The so-called "inference model" refers to a model paradigm specifically optimized for solving complex problems. Its core idea is to split the problem into multiple sub-problems that can be solved step by step, and then draw the final conclusion through multi-step reasoning. People close to the DeepSeek plan revealed that the company is expected to release V4 next week with a shorter technical description document focusing on key improvement points, and to launch a more detailed technical report about a month later to systematically disclose the model architecture and training methods.
At the same time, the controversy surrounding intellectual property and model “borrowing” is also heating up. Just earlier this week, the American AI company Anthropic accused DeepSeek and two other Chinese AI laboratories of carrying out so-called "distillation attacks" on its models, that is, using the output of more powerful models to train small models, so that the latter can approach the performance of the former without directly using the same level of computing resources. Huawei, DeepSeek and Cambrian did not respond to requests for comment.