DeepSeek Launches V4: Trillion-Parameter Multimodal Model Optimized for Chinese Hardware

DeepSeek has released V4, its fourth-generation flagship model — a trillion-parameter multimodal system capable of generating text, images, and video with context windows exceeding one million tokens. The model was released under an open-source license and is optimized for Huawei Ascend and Cambricon hardware, marking a significant step in China's effort to build a competitive AI ecosystem independent of Nvidia GPUs.

Architecture and Capabilities

DeepSeek V4 uses a mixture-of-experts architecture with approximately one trillion total parameters and an estimated 100 billion active parameters per inference pass. The model supports text, image, and video generation within a unified architecture, allowing it to reason across modalities — analyzing images and generating text descriptions, or producing video content from text prompts. The context window exceeds one million tokens, building on the expansion DeepSeek introduced to its web application in February.

Chinese Hardware Optimization

The most strategically significant aspect of V4 is its optimization for non-Nvidia hardware. DeepSeek has trained and tuned the model to run efficiently on Huawei's Ascend 910C accelerators and Cambricon's MLU370 chips — Chinese-developed AI hardware that is not subject to U.S. export controls. While Nvidia's A100 and H100 GPUs remain the de facto standard for AI training globally, DeepSeek's demonstration that frontier-competitive models can be trained on Chinese hardware challenges the assumption that export controls will durably limit China's AI capabilities.

Competitive Positioning

V4's release was timed to coincide with the opening of China's "Two Sessions" annual parliamentary meetings on March 4, ensuring maximum domestic media coverage. The model positions DeepSeek alongside Alibaba's Qwen and Baidu's Ernie as one of China's three leading AI platforms, with DeepSeek differentiating through its commitment to open-source distribution and its willingness to publish detailed technical reports about its training methodology. For the global AI community, V4 provides another data point that the gap between Chinese and Western AI capabilities continues to narrow.