DeepSeek’s latest model underwhelms

On April 24, DeepSeek finally shipped its latest model – V4. 

• Expectations were sky-high after its previous models – V3 and R1 – stunned the market and narrowed the gap with frontier US models.

Record scratch: Despite the hype, the new model landed flat.

• It clearly lags behind leading US models, and it’s not even obvious that it beats top domestic rivals (WeChat 1 and WeChat 2).

But headline performance isn’t the whole story: V4 comes with deeper integration with Chinese chips – especially Huawei Ascend – and introduces a new model architecture designed to squeeze more out of limited compute.

Get smart: Despite the disappointing performance, V4 still represents a step forward for China's AI ambitions.

• DeepSeek’s open-source research provides other Chinese labs with a playbook for working around the same compute constraints – and for switching from Nvidia chips to domestic alternatives. 

Get smarter: Constraint has driven innovation, forcing DeepSeek to extract more intelligence from limited resources, but that same constraint caps the company’s upside.

• The firm is now reportedly seeking outside funding for the first time to retain talent and buy more compute, despite long resisting external capital – which may be the only way it can stay competitive in the model race.