Baidu Releases Performance Report: Qfan Large Model Platform Reaches Nearly 10,000 Monthly Active Enterprises
-
Since the launch of Ernie Bot in March, Baidu Intelligent Cloud has now delivered its performance report on large model services after six months. On September 5, Baidu Intelligent Cloud announced that the Qfan Large Model Platform serves nearly 10,000 monthly active enterprises, covering over 400 scenarios in industries such as finance, manufacturing, energy, government affairs, and transportation. The platform has also been upgraded to version 2.0, integrating 42 mainstream large models from domestic and international sources. For third-party large models beyond Ernie Bot, the Qfan Platform not only provides simple access but also offers capabilities such as Chinese language enhancement, performance optimization, and context enrichment. Additionally, Baidu revealed plans to soon release Ernie Bot 4.0.
Notably, in terms of large model applications, Baidu Intelligent Cloud has reconstructed four industry-specific solutions based on Ernie Bot: the digital government solution—Jiuzhou, the financial solution—Kaiyuan, the industrial solution—Kaiwu, and the intelligent transportation solution—ACE 3.0.
According to reports, as large model adoption in industries is still in its early stages, many enterprises are unsure how to utilize large models effectively. To address this, Baidu Intelligent Cloud has pioneered the creation of 'industry showcases' in four key sectors: government affairs, finance, industry, and transportation.
Integrating 42 Mainstream Large Models
The Baidu Intelligent Cloud Qfan Large Model Platform, launched in March, is the world's first one-stop enterprise-level large model production platform. It not only provides large model services based on Ernie Bot or third-party open-source models but also offers a complete toolchain and development environment to help enterprises develop their own proprietary large models.
'Any new technology requires corresponding methodologies and tools to truly take root,' said Shen Dou, Executive Vice President of Baidu Group and President of Baidu Intelligent Cloud Business Group. 'Over the past six months, we have accumulated experience and tools, making significant upgrades to the Qfan Platform to make it more user-friendly and considerate.'
The upgraded Qfan Large Model Platform 2.0 integrates 42 mainstream large models from domestic and international sources, pre-installs 41 high-quality datasets with industry-specific features, and includes 10 selected application templates such as knowledge Q&A, customer service dialogue, and code assistance. These features significantly lower the barrier for enterprises to use, train, and infer large models. For third-party large models beyond Ernie Bot, the Qfan Platform provides enhancements like Chinese language optimization, performance improvements, and context enrichment. For example, foreign models like Llama2, which previously performed better in English, now deliver equally strong results in Chinese.
The upgraded toolchain of the Qfan Platform has also become more comprehensive and lightweight, covering the entire lifecycle of large models. Previously, validating a large model's performance took at least a week, but now, with the one-stop toolchain, enterprises can test a model and see results in just one day. Notably, the toolchain includes 103 high-quality Prompt templates and automated Prompt engineering capabilities, making the process more convenient and efficient.
According to reports, the Qfan Platform has also upgraded its computing services to further reduce users' computing costs and time. In large-scale cluster training with 10,000 GPUs, engineers previously spent 30%–40% of their time on fault tolerance and recovery. Now, with Baidu Intelligent Cloud's self-developed cluster networking fault management mechanism, effective model training time exceeds 95%. Leveraging distributed parallel training strategies, the Qfan Platform achieves a 95% acceleration ratio in 10,000-GPU clusters, fully unleashing the cluster's computing power. The platform is also compatible with mainstream chips and operating systems worldwide, allowing algorithms to run with minimal modification costs.
Building Industry Showcases on the Qfan Platform
'As large model adoption in industries is still in its early stages, many enterprises are unsure how to utilize large models effectively. Baidu Intelligent Cloud has pioneered the creation of 'industry showcases' in four key sectors: government affairs, finance, industry, and transportation,' Baidu stated.
This time, Baidu Intelligent Cloud introduced four industry solutions reconstructed based on Ernie Bot: the digital government solution—Jiuzhou, the financial solution—Kaiyuan, the industrial solution—Kaiwu, and the intelligent transportation solution—ACE 3.0.
Taking Jiuzhou as an example, the solution, powered by the Ernie Government Affairs Large Model, bridges the digital divide. Centered on the core principles of digital government construction—'unified governance, unified services, and unified collaboration'—it leverages large model capabilities to create a 'new' paradigm for urban governance, a 'new' experience for public and business services, and 'new momentum' for government operations.
In urban governance, city managers previously could only analyze urban operations based on a limited number of high-frequency scenarios and fixed indicators. Now, large models connect and activate various events, dynamically generating required scenarios and indicators through simple queries to identify root causes of issues. These models, trained on historical cases, past handling experiences, and accumulated knowledge, can provide real-time decision support and handling suggestions when new incidents occur.
Liu Jie, General Manager of Baidu Intelligent Cloud's Smart City division, explained with examples that urban governance previously involved two main processes: 'one-screen city monitoring' and 'networked city management.' Traditional monitoring systems could only analyze preset indicators, scenarios, and data. For instance, when digital economy became a new priority, all relevant data had to be manually collected from various departments, processed, and developed into display indicators—a tedious and time-consuming process. "With large models, this becomes much simpler," Liu noted. "Digital economy data already exists in business systems; the models can activate and integrate this data for flexible statistical analysis, making static data 'dynamic.'"
In incident handling, reliance on individual experience is being transformed. For noise complaints, 12345 hotline operators previously had to manually search knowledge bases and rely on personal expertise to route cases—a process that could be lengthy for inexperienced staff. "Large models automatically generate handling suggestions," Liu explained, "including relevant departments, routing decisions, and legal explanations for citizens, by learning from numerous past cases and converting individual experience into real-time intelligent systems."
Large models also enhance government service efficiency. In one major city, 2022 saw 1.85 million public service consultations—mostly manual, creating staffing bottlenecks. The challenge lay in translating citizens' colloquial requests into official policy terms. Now, models accurately understand requests, integrate administrative knowledge, and match queries to policies, providing faster, precise responses. This creates a personalized assistant experience, resolving issues in minutes with conversational ease.
Liu revealed high government interest in administrative large models, as they address long-standing challenges. Their integration with urban governance, public services, and office operations—bringing transformative efficiency gains—has drawn significant attention. Baidu is collaborating with multiple governments on pilot applications.
For example, Haidian's 'City Brain' is developing AI-native applications using the 'Jiuzhou' platform, aiming to connect events through simple dialogues, rapidly generate statistics, and provide handling suggestions within minutes. Deputy District Mayor Xu Zhentao stated: "We're actively exploring large model technology, from infrastructure to applications, with leading domestic models like Baidu's ERNIE participating."