According to the official introduction from Baidu Cloud, the upgrade of Qianfan 2.0 has brought multiple improvements:
In terms of toolchains, the various stages of large-scale model development, training, deployment, and application development are closely integrated. According to the introduction, Qianfan’s complete toolchain can seamlessly connect all stages of the business process throughout the entire lifecycle, maximizing efficiency. Previously, it took a technical team one month to debug and run a large model to see its effects. With Qianfan’s all-in-one toolchain platform, this can be achieved on the same day. It is worth noting that Qianfan’s platform includes 103 sets of high-quality prompt templates in its toolchain and also has automated prompt engineering capabilities, making it more convenient and easier for users.
In terms of computing power efficiency, Qianfan’s underlying system is a set of heterogeneous computing clusters, high-speed interconnected networks, and high-performance storage. It can support AI computing infrastructure for training models at a scale of tens of thousands of cards, reducing the cost of large-scale model training and inference to a minimum.
In terms of data, Qianfan not only manages large models but also enhances their capabilities. This new upgrade includes Chinese language enhancement, which significantly improves the Chinese comprehension and generation abilities of large models such as LIama2, achieving good results in Chinese conversations. At the same time, it comprehensively improves training and inference performance, with a 1/4 increase in overall throughput for training LlaMA 2 and up to a threefold improvement in inference.
In terms of enterprise-level security, with Qianfan’s security mechanisms, enterprises are able to build a four-layer security defense line to comprehensively protect the content security, data security, and system security of large-scale models. For customers with high compliance requirements, Qianfan 2.0 supports mainstream chips and operating systems both domestically and internationally. It also supports private deployment to ensure that each enterprise has independent control over its own large-scale models.
In addition, at the 2023 Baidu Cloud Intelligence Conference, Shen Dou, Executive Vice President of Baidu Group and President of Baidu Intelligent Cloud Business Group, stated that as of now, the cumulative number of users using Baidu Wangpan’s “Cloud Yiduo” intelligent assistant has exceeded 6 million.
Compared to the past, Qianfan has integrated 42 mainstream large-scale models from domestic and foreign sources. It also includes 41 high-quality industry-specific datasets and pre-installed 10 selected application paradigms such as knowledge question-answering, customer service dialogue, and code assistance. This significantly reduces the threshold for enterprises to use, train, and infer large-scale models. For third-party models other than Wenxin’s model, the Qianfan not only provides simple access but also offers Chinese enhancement, performance enhancement, and context enhancement capabilities. For example, previously foreign models like Llama2 that performed well in English conversations can now achieve the same level of performance in Chinese.
Qianfan Large Model Platform is a one-stop enterprise-level big model platform launched by Baidu, which can provide enterprises with a complete set of tools and environments for the entire process of big model development. It claims to enable enterprises to embrace big models efficiently and cost-effectively.
At the beginning of August, Qianfan completed its first round of upgrades and fully integrated 33 large models including the LLaMA2 series, and ChatGLM2, making it the platform with the most large models in China. At the same time, Qianfan launched 103 preset Prompt templates covering more than ten scenarios such as dialogue, programming, and e-commerce.
Now, Qianfan has successfully managed 42 mainstream large models from both domestic and international sources, reaching a new record high. It also includes 41 industry-specific datasets and has nearly ten thousand active enterprises. It covers over 400 business scenarios in industries such as finance, manufacturing, energy, government affairs, and transportation.
“Lots of materials, easy to use, efficient, and reliable,” Baidu Group’s Executive Vice President and President of Baidu Intelligent Cloud Business Unit, Shen Dou, used these four keywords to describe Baidu Intelligent Cloud Qianfan 2.0.