Hugging Face Announces Open-Source Model ChatGLM-6B Wins Most Popular AI Institution Award
-
Recently, Hugging Face, as a representative of the open-source AI community, summarized the top 15 most popular companies and institutions in the community, covering almost all the prominent AI technology organizations at home and abroad. Stability AI, Meta AI, and Runway ranked in the top three, while well-known names like OpenAI, Google, and Microsoft also made the list.
The only domestic institution selected was the KEG Lab, which made the list this year with its open-source model ChatGLM-6B.
When it comes to the KEG Lab, one cannot overlook its star project, ZhiPu AI, which translates its research achievements into practical applications. The highly anticipated ChatGLM-6B was jointly developed by the KEG Lab and ZhiPu AI, serving as an excellent example of industry-academia collaboration. Founded in 2019, ZhiPu AI was led by its current CEO, Zhang Peng, who steered the project out of the lab and onto the path of commercialization. Long before ChatGPT became a household name, ZhiPu AI had already ventured into the field of large models, developing the GLM pre-training architecture in 2020. The following year, it completed the training of the 10-billion-parameter model GLM-10B and successfully trained a trillion-parameter sparse model using the MoE architecture.
In 2022, ZhiPu AI and the KEG Lab jointly released the GLM-130B, a bilingual (Chinese-English) ultra-large-scale pre-training model with 130 billion parameters. In a comprehensive evaluation conducted by Stanford University's Center for Large Models that year, GLM-130B was the only Asian model among the 30 mainstream large models globally. The evaluation results showed that its key metrics, such as accuracy, were on par with or close to those of models from OpenAI, Google Brain, Microsoft, and NVIDIA. According to official reports, GLM-130B has been applied for by over 1,000 institutions in more than 70 countries worldwide. Therefore, when ZhiPu AI announced the official launch of ChatGLM and open-sourced its 6B-sized model in March this year, it had already garnered widespread attention.
It is reported that ChatGLM-6B officially launched on March 14, 2023, becoming one of China's first open-source large models. It was upgraded to ChatGLM2-6B in June and further to ChatGLM3-6B in October. Since its release, ChatGLM-6B has gained 53,000 stars and over 7,300 forks on GitHub. On Hugging Face, it has accumulated more than 10 million downloads, ranking first in the four-week trend chart. Additionally, hundreds of projects or applications based on ChatGLM-6B have emerged, such as LoRA, P-Tuning, and langchain.