DeepSeek to launch V4 with native multimodal support

The Financial Times reported on February 28 that DeepSeek, a China-based AI developer, would release V4 next week with native support for image, video, and text generation.

Sources cited by FT said DeepSeek partnered with Huawei and Cambricon to optimize V4 for Chinese domestic chips and would prioritize early access for domestic suppliers.

Those sources said DeepSeek did not share preview builds with Nvidia or AMD, which broke industry-standard practices, though DeepSeek had previously collaborated with Nvidia technical personnel.

A preview build codenamed “sealion-lite” was reportedly in closed internal testing, with an expanded context window of about 1 million tokens and native multimodal reasoning.

Some reports suggested V4 might be around a trillion parameters and could boost domestic semiconductor demand.

🔗 Source: Pandaily

🧠 Food for thought

Implications, context, and why it matters.

DeepSeek is shifting to domestic chips under U.S. export controls and limited supply

DeepSeek’s reported work with Huawei and Cambricon on V4 stems from constrained access to advanced compute after U.S. export controls tightened; SemiAnalysis said training on Huawei chips delayed DeepSeek’s next-generation model 1.
In response, DeepSeek has pushed compute efficiency, adding DeepSeek Sparse Attention (DSA) to cut long-context compute plus memory needs 2.
V3.2-Exp shipped with “Day 0” support from Chinese hardware firms, which set up V4 as a bigger step toward first-class support for domestic accelerators (specialized chips used to speed up AI workloads) 3.

A lower-cost AI stack is taking shape with global spillover

A coordinated V4 release would strengthen a non-NVIDIA AI stack in China, pairing domestic chips like Huawei’s with software such as Huawei’s CANN (its AI software platform for running and optimizing models on Huawei chips) 4.
Price cuts have helped adoption; after DeepSeek lowered API pricing, including a further 50%+ discount with V3.2-Exp, token usage for Chinese models exceeded American models on OpenRouter (a marketplace that routes user requests across many AI model providers) for the first time, according to a market note citing OpenRouter data 2 5.
Interest around V4, including reported V4 Lite testing, has been tied to demand plus share-price moves among Chinese cloud and hardware suppliers 5.