DeepSeek V4 AI model launches, built for Huawei chips

2 min readUpdated April 24, 2026

More than a year after its R1 reasoning model sent shockwaves through global tech markets, Chinese AI startup DeepSeek has unveiled a preview of its V4 model — the company's most significant product debut since that viral release. The new model is open-sourced and adapted to run on Huawei's Ascend AI processors, a shift away from the company's prior reliance on Nvidia $NVDA chips.

The V4 comes in two versions: DeepSeek-V4-Pro, with 1.6 trillion total parameters and 49 billion active parameters, and DeepSeek-V4-Flash, with 284 billion total parameters and 13 billion active parameters. Both support a context length of one million tokens. DeepSeek said the Pro version outperforms other open-source models in world-knowledge benchmarks, trailing only Google $GOOGL's closed-source Gemini-3.1-Pro.

JPMorgan Chase is pledging $750 billion to boost U.S. housing supply and homeownership

Agent-based tasks, which require substantially more computational resources than typical chatbot exchanges, are a primary use case DeepSeek is targeting with V4; the company noted that the model works with tools including Anthropic's Claude Code. The company also highlighted V4's performance in coding benchmarks and its efficiency in long-context processing.

Both models use a mixture-of-experts architecture and were pre-trained on more than 32 trillion tokens, according to DeepSeek's technical documentation. The models are available under an MIT license, and the API is live. DeepSeek did not provide a timeline for a final, non-preview release.

The U.S.-China AI performance gap has effectively closed in recent months, with leading American and Chinese models separated by only a few percentage points on major benchmarks. When DeepSeek disclosed that R1 was built for under $6 million on lower-capacity Nvidia $NVDA hardware, the claim unsettled markets and prompted hard questions about whether the enormous sums American tech giants were pouring into AI infrastructure were truly necessary.

Rivals including Alibaba and ByteDance have moved aggressively into the space since R1's debut, eroding DeepSeek's early dominance in China's AI market. Hong Kong markets reflected the competitive anxiety V4 generated: shares of Zhipu AI and MiniMax dropped roughly 8% to 9%, while domestic chipmakers saw a surge of investor enthusiasm — Hua Hong Semiconductor climbed 15% and SMIC advanced approximately 9% to 10%.

Nvidia CEO Jensen Huang flagged the risk of losing ground in China earlier this month. Speaking on the Dwarkesh Podcast earlier this month, Huang put the stakes plainly: "The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation," according to Reuters.

DeepSeek is back with a new open-source AI model built on Chinese chips

Related Content

Related Content