Huawei Officially Releases Pangu Model 3.0, Exploring AI Empowerment Across Industries
-
Large models are undoubtedly the hottest topic this year. General-purpose large models represent the 'brain' of human intelligence, while industrial large models serve as the 'hands and feet.' To address the inefficiencies in current AI development under the 'small workshop model,' Huawei's Pangu Model will introduce a series of solutions. On July 7, Zhang Ping'an, Executive Director of Huawei and CEO of Huawei Cloud, announced at the Huawei Developer Conference the official release of Huawei Cloud's Pangu Model 3.0.
It is understood that after the upgrade to Pangu Industry Model 3.0, 'Pangu' will serve as the umbrella term for its 'large model series,' encompassing both foundational models like language and vision models, as well as industry-specific models such as finance, manufacturing, and drug molecule models, along with scenario development services. Pangu Model 3.0 is structured into three layers: L0 foundational models, L1 industry models, and L2 scenario models, covering nearly all mainstream directions in AI today. Its core mission is to empower various industries.
'Pangu Model doesn't write poetry; it gets things done,' Zhang Ping'an reiterated Huawei's positioning in the large model field. Huawei's Pangu Model has already deeply engaged with over 10 industries, including finance, covering more than 400 business scenarios.
Innovation in large models isn't just about the models themselves but requires full-stack innovation, optimizing computing power, operators, frameworks, and platforms. Huawei revealed that through continuous R&D, Pangu Model has been upgraded in performance, depth, architecture, and data enhancement. The model leverages Kunpeng and Ascend-based AI computing cloud platforms, along with the heterogeneous computing architecture CANN, the full-scenario AI framework MindSpore, and the AI development pipeline ModelArts, achieving extreme efficiency and establishing itself as a global AI leader.
At the World Artificial Intelligence Conference - Ascend AI Industry Summit hosted by Huawei on July 6, a large model joint innovation initiative was launched. Twenty-six leading enterprises and research institutions will collaborate with Huawei on foundational and industry-specific model innovations based on Ascend AI. Additionally, Huawei and its partners jointly released the Ascend AI Large Model Training-Inference Integrated Solution to accelerate model deployment across industries. Twenty-three Ascend AI partners also unveiled new AI servers, intelligent edge, and terminal products, collectively providing rich solutions for industry upgrades.
It is reported that Ascend AI clusters now support AI computing centers in 25 cities across China, with seven city-level public computing platforms selected for the first batch of national 'Next-Generation AI Public Computing Open Innovation Platforms.' At the forum, Huawei announced a comprehensive upgrade of Ascend AI clusters, scaling from an initial 4,000-card cluster to 16,000 cards—the industry's first 10,000-card AI cluster, offering faster training speeds and stable training cycles exceeding 30 days, ten times ahead of the industry. Currently, over 30 large models have been natively developed or adapted based on Ascend AI computing power, accounting for nearly half of China's large models.
Amid the rapid advancement of large models, enterprises face the challenge of practical application. Ken Hu, Huawei's Rotating Chairman, stated at the 2023 World AI Conference that Huawei will drive systematic innovation, uphold open-source principles, and deepen industry engagement to accelerate the growth of the Ascend AI ecosystem. By collaborating with partners, Huawei aims to advance AI from theory to practice, supporting high-quality AI development with robust computing power.