[DigitalToday reporter Chi-gyu Hwang] Chinese AI developer DeepSeek said it has released preview versions of its new V4 AI model, the Wall Street Journal reported on Thursday.

DeepSeek announced the launch through its official WeChat account.

According to the company, V4-Pro has significantly improved agent coding capabilities compared with earlier models. In user experience, it outperforms Anthropic's Claude Sonnet 4.5, and its output quality is close to Claude Opus 4.6 in non-thought mode. It does not match Opus 4.6 in thought mode.

The V4 launch comes after DeepSeek repeatedly delayed major model updates while rival firms in China rolled out updates one after another. Moonshot AI, Minimax, Alibaba and ByteDance pursued updates aggressively.

The V4 model is also drawing attention for supporting Huawei chips. DeepSeek verified its V4 efficiency technology on both Nvidia GPUs and Huawei Ascend NPUs. Huawei said its full Ascend lineup fully supports DeepSeek's V4 models.

The V4 model applies Sparse Attention technology unveiled last year. Rather than processing everything at once, Sparse Attention focuses only on the most relevant parts, enabling it to handle much longer documents.

Its pricing competitiveness is also in the spotlight. The cost for 1 million output tokens on V4-Pro is $3.48, far lower than Anthropic Opus 4.6 at $25. DeepSeek also launched V4-Flash, which it described as cheaper and faster. It delivers performance comparable to V4-Pro for simple tasks but falls short of V4-Pro on high-difficulty tasks.