As the release date approaches, more technical details about the domestic AI leader DeepSeek V4 have come to light. According to a well-known leaker, Legit, DeepSeek has begun closed-door internal testing of its preview version codenamed "sealion-lite (lightweight sealion)" and all participants have signed strict non-disclosure agreements.

Compared to the existing V3 model, DeepSeek V4 will achieve a generational leap. The new model not only has native multimodal processing capabilities but also directly increases the context window to 1 million tokens. This means it can "digest" several books or long code repositories at once and perform in-depth logical reasoning. Early test feedback shows that V4 performs impressively on high-difficulty tasks such as generating complex SVG vector graphics, with performance significantly better than current online models on web and app platforms.

AIbase learned that DeepSeek V4 is expected to be a massive foundational model at the level of trillion parameters. Due to the exponential growth in model size, the training period has been extended, leading to a delay in its release compared to the original plan. Notably, DeepSeek broke industry conventions before the release and did not provide previews to U.S.-based chip manufacturers such as NVIDIA and AMD, but instead opened access weeks in advance to Chinese chip suppliers including Huawei, ensuring the model is deeply adapted and optimized on domestic computing platforms.

Meanwhile, DeepSeek's GitHub repository has recently seen frequent updates. Although much of the content is routine code cleanup after the engineering holiday, it also includes many enterprise-level solutions related to API integration. As the rumor of "release next week" intensifies, global developers are holding their breath, watching how this Chinese company, known for its "high cost-performance and high performance," will once again challenge the global AI competition landscape through DeepSeek V4.

Summary:

  • 🚀 Leap in Parameter Scale: DeepSeek V4 is positioned as a trillion-parameter model, supporting input of up to millions of text tokens, and its native multimodal capabilities will bring stronger logical and creative performance.

  • 🇨🇳 Prioritizing Domestic Computing Power: The company has shifted its focus to adapt to Chinese chip suppliers such as Huawei, aiming to reduce reliance on external supply chains and conduct in-depth performance optimization for domestic hardware.

  • 🗓️ Countdown to Launch: The lightweight version's internal testing has quietly started, and foreign media predict the official announcement could come as early as next week, a move that has already drawn significant attention from the global AI community and capital markets.