DeepSeek released a preview of V4 on Friday, its new flagship model. The open-source system matches leading closed-source rivals on performance and is the firm's first release optimised for Huawei's Ascend chips.

The model handles significantly longer prompts than its predecessor, thanks to a redesigned architecture for processing large volumes of text more efficiently. DeepSeek says performance is on par with closed-source offerings from Anthropic, OpenAI, and Google.

The Huawei detail matters most. V4 is a direct test of whether Chinese AI labs can reduce their dependence on Nvidia hardware, which remains subject to US export controls. If the Ascend-optimised version holds up in production workloads, it weakens one of Washington's primary pressure points on China's AI sector.

DeepSeek has built its reputation on doing more with less. Its earlier V3 and R1 models drew attention for training efficiency that undercut Western labs on cost. V4 continues that pattern whilst expanding capability.

The model remains open-source, consistent with DeepSeek's approach since its founding. That positions it as a direct alternative to closed offerings from frontier labs charging per-token API fees.

What this signals

The release lands in a week already dominated by AI funding and geopolitics. Google announced an investment of up to $40 billion in Anthropic at a $350 billion valuation. Meanwhile, China blocked Meta's $2 billion acquisition of AI startup Manus on national security grounds.

Taken together, the picture is one of accelerating bifurcation. US labs are consolidating around massive capital raises and closed models. Chinese labs are pushing open-weight alternatives on domestic hardware. The two tracks are diverging faster than most observers expected twelve months ago.

V4 also arrives as the AI compute crunch begins to bite the broader economy. Reports this week flagged rising electricity prices and constrained GPU availability. An efficient open model that runs on non-Nvidia silicon offers at least a partial pressure valve.

For researchers and developers outside the major labs, V4 expands the set of capable models available without API fees or usage restrictions. Whether it holds up under rigorous independent benchmarking remains to be seen. DeepSeek has not yet published a full technical report.

The firm has said a complete release will follow the preview period, though it has not given a specific date.