China’s AI landscape is buzzing again as DeepSeek ends months of anticipation, unveiling preview versions of its groundbreaking DeepSeek V4 AI model. This release marks a significant milestone, especially given the aggressive updates from domestic rivals like Moonshot AI and Alibaba. The flagship V4-Pro model demonstrates dramatically improved agent abilities, with internal feedback positioning it as the “go-to agentic coding model.”

It reportedly outperforms Anthropic’s Sonnet 4.5 in user experience and output quality, nearing Opus 4.6’s non-thought mode capabilities. Beyond its impressive performance, the DeepSeek V4 AI model is a testament to advanced hardware integration. DeepSeek has successfully validated its key efficiency techniques on both Nvidia GPUs and Huawei’s Ascend NPUs, underscoring Huawei’s full-stack support. The model also leverages “Sparse Attention” for handling much longer documents efficiently. While V4-Pro is more expensive than previous DeepSeek versions, its $3.48 per million output tokens dramatically undercuts Western competitors like Anthropic’s Opus 4.6 at $25.

A faster, cheaper V4-Flash version is also available. As DeepSeek seeks its first external fundraising, the performance of the DeepSeek V4 AI model will undoubtedly be a key factor in its valuation.