DeepSeek V4 is a new AI model that needs to be tuned for Chinese hardware accelerators.

DeepSeek is preparing for the release of its new V4 model

This week, according to *Financial Times*, Chinese company DeepSeek plans to unveil the long‑awaited multimodal AI model — V4. A key feature will be optimization for Huawei and Cambricon accelerators, enabling Chinese firms to promote their own solutions without heavy reliance on imported technologies.

Key dates and context

- The model announcement will coincide with a parliamentary session in China that starts March 4.
- This will be DeepSeek’s biggest release since January of last year, when the company launched the R1 model. At that time it was claimed that the new system matched Western counterparts in performance while costing significantly less.

Issues and controversies

It was previously revealed that DeepSeek used data from American models to train its systems and even relied on Nvidia accelerators. However, those devices are sanctioned in China. As a result, the company deliberately did not optimize V4 for Nvidia hardware.

Early attempts to train the model on Huawei accelerators were less successful, but now DeepSeek intends to improve compatibility with these platforms.

Market impact

Optimizing V4 for Huawei and Cambricon should stimulate demand for these hardware solutions in China and reduce import dependence. Experts predict this will strengthen the position of the Chinese AI market.

Additional information

An abstract of the V4 model will be released later this week in a shortened form; a full description is expected in about a month. Meanwhile, American company Anthropic accused DeepSeek of “distilling” its own AI models, adding tension to international relations regarding artificial intelligence technologies.