(Yicai) April 24 -- China’s DeepSeek released a preview of its much-anticipated new flagship artificial intelligence model within hours of OpenAI rolling out GPT-5.5, highlighting the accelerating pace of the global AI race.

The Hangzhou-based startup announced the preview version of DeepSeek-V4 around noon today. It runs on Chinese computing infrastructure through a partnership with Huawei Technologies and its Ascend platform, reflecting the country’s push to reduce its reliance on foreign chips.

DeepSeek vaulted onto the global stage in January last year with the release of a ChatGPT-style model, which paired strong reasoning performance while slashing operating costs. Silicon Valley venture capitalist Marc Andreessen famously described it as an “AI Sputnik moment.”

DeepSeek-V4 comes in two versions -- Pro and Flash -- aligned with the expert and fast modes on DeepSeek’s website and app. The Pro version has 1.6 trillion parameters with 49 billion activated parameters and is trained on 33 trillion tokens of data, while the Flash version has 284 billion parameters with 13 billion activated and 32 trillion tokens of pre-training data.

The model supports a context window of one million tokens and has achieved leading performance in China and among open-source models in agentic capabilities, world knowledge, and reasoning, according to media reports.

“From now on, a one-million-token context window will be a standard feature for all official services of DeepSeek,” the firm said, adding that V4 introduces a new attention mechanism that compresses tokens and, combined with DeepSeek Sparse Attention, significantly reduces computing and memory requirements compared with traditional methods.

In terms of pricing, DeepSeek said the input cost for V4-Pro is CNY1 (15 US cents) per million tokens and the output cost is CNY12 (USD1.80), while V4-Flash costs CNY0.2 per million tokens for input and CNY2 for output.

However, the company noted that the Pro version currently has limited service throughput due to constraints in high-end computing power. It expects prices to drop after Huawei’s Ascend-based Atlas 950 SuperPoD -- a high-performance, liquid-cooled AI computing cluster -- is deployed at scale in the second half of the year.

Shortly before the launch, media reports said DeepSeek had kicked off its first external fundraiser. Additional capital is expected to help the firm secure more computing resources, accelerate model development, and offer more competitive compensation to retain top talent. 

The DeepSeek-V4 release does not include a multimodal version, leading to speculation that constraints in computing power and funding may have delayed its development, even as multimodal capabilities become standard among leading model providers.

Editor: Emmi Laine