Tmtpost, a Chinese online tech publication, reported today that Huawei Technologies will release a multimodal large language model (LLM) that is benchmarked against ChatGPT, named Pangu Chat.
Huawei is among a number of Chinese tech giants that have announced their ambitions to develop a LLM in recent months. The telecom equipment titan started its foray into LLM as early as November 2020, and has worked on it ever since.
After OpenAI’s ChatGPT took the global tech community by storm, Huawei began to reveal the headway it has made in developing a LLM it calls “Pangu.”
According to Tmtpost, Huawei is set to unveil Pangu Chat at HDC.Cloud 2023 on July 7 this year, as well as open an API for internal tests of the chatbot service. Pangu will target clients in the business and government sectors.
Huawei published data earlier showing that PanGu-Σ — the latest edition of the LLM — contains as many as 1.085 trillion parameters and is built upon Huawei’s proprietary MindSpore framework.
Overall, Huawei said Pangu-Σ could rival GPT-3.5 in conversational ability.
Public records show that the Pangu LLM made its debut in April 2021, and received an upgrade to the 2.0 version in April the following year. As time went on, more parameters and functions were being added to the model to iterate it.
As of now, Huawei’s Pangu has become an all-in-one model, with natural language processing (NLP), computer vision (CV) and climate modelling all tagged as “would-be” add-ons to be included in the existing LLM.
Take meteorological reports, a key use case for Pangu. Huawei claims that its key metrics and accuracy both surpass the most advanced weather forecast systems, while improving the speed by more than 1,000 times.
At the same time, Pangu’s climate model supports a wide range of forecast tasks, such as predicting the trail of typhoon. Compared to the methodology behind traditional weather reports, using the Pangu LLM can lower the margin of error by more than 20%, said Huawei.
The tech titan has said that its LLM will stand out for three aspects, including the capacity to absorb massive data, a robust network structure, and sufficient versatility in terms of application to be called a “general-purpose” model.
Aside from NLP, CV and climate modelling, which constitute the bottom layer of Pangu, the top layer of the model will be a result of collaborations between Huawei and industry partners, Tmtpost said, citing slides shown by Huawei Cloud executives on previous occasions.