DeepSeek V4 Goes China-First on Domestic Chips

DeepSeek’s V4 multimodal model is going all-in on China-made hardware, a move that could reshape the domestic AI stack and ripple through global sourcing decisions.

DeepSeek is lining up a major update to its multimodal large language model, with native image, video, and text generation capabilities. The next-week release, according to a Financial Times blog post summarized by Pandaily, marks the company’s first substantive version jump since the R1 reasoning model debuted in January 2025. The V4 preview codename is “sealion-lite,” and a two-stage rollout is planned: a concise technical note at launch, followed by a fuller engineering report about a month later.

Crucially, DeepSeek is tying V4 tightly to domestic compute power. The company has reportedly partnered with Huawei and Cambricon to optimize hardware adaptation for the model, a move designed to accelerate inference on Chinese chips and to bolster demand for homegrown semiconductor capacity. In a departure from industry practice, the firm is prioritizing early access for domestic suppliers and has not provided test versions to NVIDIA or AMD, a decision that effectively shifts the model’s testing and optimization away from the familiar Western hardware ecosystem.

This shift sits inside a broader policy and market context. The move aligns with a Chinese emphasis on “自主可控” (self-reliant control) of critical AI infrastructure and the push to strengthen domestic compute supply chains as AI models grow in scale and cost. If V4’s domestic-hardware emphasis proves durable, it could tilt early performance benchmarks toward China-made chips and compilers, while raising the bar for international players to demonstrate cross-platform efficiency across a broader hardware stack.

From a supply-chain vantage point, the decision to forego pre-release access to NVIDIA/AMD signals a deliberate domestic-focus strategy. DeepSeek has long collaborated with NVIDIA on prior stacks, but the new approach suggests a longer-term bet on Huawei, Cambricon, and other homegrown suppliers to provide both inference accelerate and ecosystem alignment. That could translate into faster, cheaper deployment for Chinese customers and a more self-contained AI economy, but it also raises questions about portability, external validation, and vendor-lock risk for multinational users who need multi-hardware compatibility.

For global manufacturers and investors, the development is a reminder that China’s AI ecosystem is not merely a code-and-cloud story; it’s becoming a hardware-and-regulation story as well. If DeepSeek’s V4 gains traction on domestic chips, enterprise buyers may face a more complex procurement calculus: cost and performance must be weighed against potential vendor-lock and the pace of domestic chipset maturity. The engineering delays, certification cycles, and optimization gaps that often accompany a hardware-specific stack could become a more visible part of deployment timelines.

Two to four practitioner takeaways emerge clearly. First, the hardware path matters more than ever. For teams planning deployments in China, domestic compute ecosystems—especially Huawei’s and Cambricon’s—will be the default consideration, not an afterthought. Second, this hints at higher leverage for Chinese chip-makers in AI inference, which could shift capital toward domestic accelerators and away from imported GPUs. Third, expect a tighter coupling between model software and hardware optimization; portability across platforms may lag unless DeepSeek or rivals commit to cross-hardware standards. Fourth, until the technical white paper lands, performance and cost metrics remain uncertain; buyers should watch for benchmarks and third-party validations once the Sea-lion-lite test phase concludes.

The next steps seem straightforward: await DeepSeek’s technical explainer and, a month later, a deeper engineering report that will illuminate how the V4 stack performs on Huawei-Cambricon hardware and what that implies for cross-border deployment. In the meantime, the industry should prepare for a new pattern in China’s AI race—one where the model’s fate is increasingly tethered to the trajectory of domestic compute and the policy environment that prioritizes homegrown capability.

Sources

DeepSeek to Release V4 Multimodal Modelwith Native Image, Video, and Text Generation Support