China’s DeepSeek launches long-awaited AI model

The company said on its official WeChat account on Friday that the V4-Pro’s agency capabilities have been significantly improved compared to previous models. They note that the model is now the “preferred surrogate coding model” internally, with feedback showing it beats Anthropic’s Sonnet 4.5 in terms of user experience and delivers output quality closer to Opus 4.6’s non-thinking mode, although it still lags behind Opus 4.6’s thinking mode.

This release ends the long wait for a major update to the DeepSeek model. While the company has repeatedly delayed releasing major model updates, domestic rivals including Moonshot AI’s Kimi, MiniMax, Alibaba Group and ByteDance are actively rolling out updates.

The Chinese company’s V4 model also marks a milestone for domestic chips.

DeepSeek said it validated one of V4’s key efficiency technologies on Nvidia GPUs and Huawei’s Ascend NPUs. Huawei said in a WeChat post that its entire Ascend series now offers full-stack support for DeepSeek V4 models.

The V4 model also uses “sparse attention” technology introduced last year, which allows the model to focus on only the most relevant parts rather than processing everything at once. The company says this enables the model to handle longer documents.

While the V4-Pro is significantly more expensive than DeepSeek’s previous models, it’s still significantly cheaper than its Western rivals. For example, Anthropic charges $25 per million output tokens for its Opus 4.6 model, while V4-Pro charges $3.48 for 1 million output tokens.

DeepSeek also introduced V4-Flash, a cheaper and faster version that matched V4-Pro on simpler tasks but fell behind on more demanding tasks.

The startup is seeking at least $300 million in first outside funding, with investors saying the company’s valuation will be tied to the performance of its latest models, The Wall Street Journal reported.

Write to Tracy Qu: tracy.qu@wsj.com