Huawei, the Chinese tech super-giant, is planning to launch “PanGu Chat,” a new multimodal large-scale model that directly competes with OpenAI’s ChatGPT. The official launch and internal testing of PanGu Chat are scheduled for July 7th at the Huawei Cloud Developer Conference. This model targets enterprise customers in the B2B and B2G sectors. PanGu-Σ model is developed within Huawei’s MindSpore framework and boasts an impressive 1.085 trillion parameters.
It is an extension of Huawei’s PanGu model, which was introduced in 2021 and upgraded to version 2.0 in 2022. The PanGu model includes modules for NLP, CV, multimodal, and scientific computing models.
The PanGu NLP model utilizes an Encoder-Decoder architecture, combining understanding and generation capabilities for flexibility in different systems. It can be fine-tuned with minimal data and learnable parameters, making it suitable for quick deployment in various applications. Huawei, known for its innovative smartphones, is gearing up to launch its own AI chatbot to compete with ChatGPT and other popular AI chatbots in the market. Huawei’s ChatGPT rival, called Pangu Chat, is expected to be launched by July 7 at the Huawei Developer Conference (Cloud) event. The PanGu CV model is the industry’s largest and supports on-demand extraction, excelling in intelligent inspection and smart logistics.
It has improved small-sample learning capabilities and leading performance in the industry.
Huawei’s PanGu meteorological model offers real-time weather forecasting with high accuracy, utilizing innovative network structures and algorithms. Training the PanGu model involves over 2,000 Ascend 910 chips and takes more than two months. Huawei has invested substantial resources, including 4,000 GPU/TPU cards and a total cost of 960 million Chinese yuan over three years, to support the model’s development. Experts predict that Huawei’s PanGu model will establish itself as a leading model in China, benefiting Huawei’s ecosystem companies and empowering various industries.
Recent reports suggest that Pangu Chat will initially be available for business customers and may be sold to the government and enterprise sectors. Pangu Chat is based on the large-scale Pangu cloud model, which has 100 billion parameters, lower than OpenAI’s GPT-3 but comparable to GPT-4. Parameters define the skill of the model, and while Pangu Chat has 100 billion parameters, it may not match the capabilities of OpenAI’s GPT-4. Huawei utilized over 2,000 Ascend 910 chipsets and 4,000 GPU/TPU cards for training the PanGu model, with a computing power cost of over 960 million yuan in three years.
The Pangu Series AI Large Models consist of NLP, CV, and scientific computing models, with the uppermost layer developed with partners.