Generative AI Models: DeepSeek Launches New V4 Preview in China

DeepSeek launched preview versions of its latest generative AI models, the V4 series, featuring 1 million token context windows. The Silicon Review reports on the Pro and Flash versions, agentic coding capabilities, and China's push for AI self-reliance.

DeepSeek, the Chinese artificial intelligence startup that shook global markets with its R1 model in 2025, released preview versions of its latest generative AI models on Friday as the technology rivalry between China and the United States intensifies. The DeepSeek-V4 series introduces two versions: Pro for high-performance tasks and Flash for cost-effective deployment, both featuring a breakthrough 1 million token context window.

The 1 million token context lengths represent a dramatic upgrade from V3's 128,000 tokens. To put this in perspective, the model can process the entire "Three-Body Problem" trilogy all three books in a single pass. DeepSeek achieved this through a novel token-dimension compression mechanism combined with DSA sparse attention technology, which dramatically reduces computational and memory requirements compared to traditional methods.

Agentic capabilities represent the most significant leap from previous versions. The V4-Pro version has become the default agentic coding model for DeepSeek's internal engineering team. Internal evaluations show it outperforms Claude Sonnet 4.5 and delivers quality approaching Claude Opus 4.6's non-reasoning mode, though it still lags behind on complex reasoning tasks. The V4-Flash version matches the Pro version on simpler agent tasks and offers comparable reasoning for everyday queries.

On knowledge benchmarks, V4-Pro leads all open-weight models and trails only the proprietary Gemini-Pro-3.1. In mathematics, science, and competitive coding tests, it surpasses every open-source model with published results. The model was trained on an unprecedented 14.8 trillion tokens.

DeepSeek positions V4 as an open-source "bridge" to self-reliance, particularly given US export controls on advanced semiconductors. The models reportedly run on Huawei's latest Ascend chips, requiring a complete rewrite of core code to migrate from Nvidia's CUDA ecosystem to Huawei's CANN architecture a strategic move that signals deepening technological decoupling between the world's two largest economies.

DeepSeek announced that 1 million token context windows will become standard for all its official services going forward. Pricing remains unchanged: Pro is $2.19 per million input tokens and $7.19 per million output tokens, while Flash is $0.27 per million input tokens and $1.07 per million output tokens.

As DeepSeek launches V4 with 1 million token context and competitive agentic coding performance, The Silicon Review examines how China's flagship generative AI models are closing the gap with Western frontier models and why hardware independence is as important as benchmark scores in the new AI arms race.

About the Author

Sashindra Suresh is an experienced writer specializing in artificial intelligence, software development, and emerging technologies. With a strong ability to translate complex technical concepts into clear, engaging insights, she has contributed to a wide range of publications and platforms. Her work focuses on making cutting-edge innovations accessible to both industry professionals and curious readers alike.