According to the information, Huawei could launch this software as “Pangu Chat”, a name derived from its large-scale learning model. Also, it’s expected that this AI tech could unveil at Huawei Developer Conference (Cloud) on July 7th, 2023.
However, more details suggest that Huawei will only make it available for business purposes and sell it to government and enterprise customers.
It’s reported that Huawei has tested and verified the work of Pangu Chat using a large-scale Pangu cloud model with 100 billion parameters.
For testing, the company used over 2000 Huawei Ascend 910 chipsets for over 2 months of data training capabilities. According to industry analysts, more than 4000 GPU/TPU cards are used in large-scale model training each year, and the computing power cost of large-scale models in three years has reached high as 960 million yuan.
Aside from tests, the specs of the Huawei Pangu model reveal that the new Pangu-E large model could output at least 1.085 trillion parameters, which is equal to the capability of GPT 4. This model is developed by Huawei using the MindSpore framework. And it can generate dialogues to the level of GTP-3.5.
It was officially launched in 2021 and version 2.0 came in April 2022. Currently, the Natural language processing (NLP) large model, CV large model, and Meteorological model are empowering this platform.
Specifically, the basic layer of Huawei’s “Pangu Series AI Large Models” mainly includes NLP large models, CV large models, scientific computing large models, and other learning models. Meanwhile, the upper layer consists of Huawei’s large models developed with partners.
Pangu is the first Chinese pre-training large model with over 100 billion parameters and the CV large model has reached 3 billion parameters.
It’s the first to achieve both discrimination and generation data capabilities. It also has the ability to also learn small samples, second-level weather forecasts, and map, text, and audio three-modal large-scale learning.
Pangu NLP large model uses the Encoder-Decoder architecture taking into account the comprehension and generation capabilities of the NLP large model.
It also ensures the flexibility of embedding the model in different systems. In downstream applications, only a small number of samples and learnable parameters are needed to complete the rapid fine-tuning and downstream adaptation of a large-scale model of 100 billion.
Huawei has not announced this news. Despite no confirmation, it’s good to see that Huawei is using AI power to deliver better solutions, and launching a ChatGPT rival would be a cherry on top.
However, it would be more interesting if it could make it available on Huawei smartphones.