Huawei chips refine DeepSeek model in major leap for China's AI self-reliance (www.scmp.com)

Huawei has achieved a significant milestone in advancing China's AI self-reliance by refining its DeepSeek model, which now features 1.6 trillion parameters. This breakthrough was made possible through a computing cluster powered by at least 1,000 Huawei chips, allowing for "full-parameter" post-training—a complex process previously dominated by overseas technology. This new capability enables the model to not only understand language but also effectively execute tasks by following human instructions and safety protocols. The implications of this advancement are profound for the AI/ML community in China. While local chipmakers have excelled in AI inference, they have struggled with training, which is essential for creating versatile and sophisticated AI models. The successful integration of full post-training marks a pivotal shift—it transforms a linear input-output process into a more dynamic system that can self-reflect and adapt. This level of technological sophistication is expected to enhance the autonomy of China's AI industry chain, positioning the country as a more formidable player in the global AI landscape.