Cambricon Achieves "Day 0" Compatibility for DeepSeek-V4 to Bolster Domestic AI Ecosystem

Cambricon, a leader in AI hardware acceleration, has announced full "Day 0" compatibility for the newly released DeepSeek-V4 model series. This immediate support ensures that both the 285-billion parameter variant and the ultra-large 1.6-trillion parameter MoE (Mixture of Experts) version can run with peak performance on Cambricon’s latest NPU architectures from the moment of launch.

To support the broader developer community, Cambricon has also open-sourced its optimized kernel code on GitHub. These optimizations are specifically designed to maximize memory bandwidth and reduce latency during the inference phase, which is critical for real-time Agentic applications. This move is seen as a major step toward building a self-reliant domestic AI stack that can compete with international hardware providers.

By ensuring that top-tier open-source models like DeepSeek-V4 are immediately usable on domestic chips, Cambricon is helping to reduce the "Software Gap" that has historically hindered the adoption of non-Western hardware in high-end AI research. "Our goal is to ensure that great models have great hardware to run on, regardless of global supply chain shifts," stated a Cambricon engineering lead.