#What is DeepSeek's New V4 AI Model?
DeepSeek recently introduced the preview version of its V4 AI model. This new open-source system features a remarkable capability of handling a 1 million token context window. Additionally, it supports Huawei’s domestic AI chips.
DeepSeek offers two distinct versions of the V4 model. The V4 Pro model boasts an impressive 1.6 trillion total parameters along with 49 billion active parameters. In contrast, the V4 Flash model has 284 billion total parameters and 13 billion active parameters. These specifications highlight the diversity in performance that each model can provide.
#How Do the Two Versions of V4 Compare?
The V4 Pro version is crafted for high-performance scenarios, excelling in complex tasks such as agentic coding and competitive programming. Conversely, the V4 Flash is specifically designed to deliver faster and more cost-effective solutions. However, it is important to note that despite its advanced capabilities, the V4 Pro still falls short compared to leading closed-source models from industry giants like Google and OpenAI.
#What Does This Launch Mean for AI Development in China?
This launch marks a significant milestone for DeepSeek, being the first major model tailored for Huawei hardware. With the release of V4, Huawei’s Ascend supernode equipped with Ascend 950 AI chips will provide comprehensive support for this new model. This shift signifies China's intention to reduce its dependency on foreign AI infrastructure, a necessity that has arisen due to recent U.S. export restrictions. Previously, DeepSeek relied on Nvidia chips for its development, but the latest models incorporate domestic alternatives incorporating Huawei’s infrastructure to enhance the capabilities and training of the Flash version.
#What Impact Does DeepSeek Aim to Achieve?
The launch of the V4 model comes more than a year after the introduction of DeepSeek’s R1 model, which significantly influenced global AI markets by showcasing that a Chinese lab could construct competitive models at lower costs than previously perceived. The V4 model is being positioned not just as another release, but as a formidable challenge to prominent U.S. AI firms, claiming to compete effectively while maintaining an open-source framework for public access.
To facilitate user engagement, DeepSeek has made the V4 model accessible through chat.deepseek.com in both Expert Mode and Instant Mode. An update to its API was also implemented on the same day of the release, signaling a proactive approach to further integration and usability.