Baidu Founder and CEO Robin Li

TMTPOST -- AI models and chips have no real value without meaningful applications, said Baidu founder and CEO Robin Li at the company's Create Developer Conference in Wuhan on Sunday.

Li officially launched Baidu’s latest offerings — the Wenxin Large Model 4.5 Turbo and Wenxin Large Model X1 Turbo — aimed at addressing common industry pain points such as slow speeds, high costs, limited modalities, and persistent hallucination rates.

“DeepSeek has made breakthroughs, and protocols like MCP are becoming industry standards. Yet, developers remain anxious,” Li said, noting the rapid pace of model upgrades. “They fear that the applications they build today will soon be rendered obsolete.”

Li highlighted that in just the past two quarters, over 100 large model updates and launches have flooded the market, intensifying competition among vendors but leaving developers hesitant. He argued that while faster, more capable models offer new opportunities, they also demand a strategic approach to selecting foundational models and identifying application scenarios.

“AI applications are the true creators of value,” Li said. “Without applications, models and chips — no matter how advanced — are essentially meaningless.”

The comments come as Baidu intensifies its push into AI. The Wenxin 4.5 Turbo claims input costs of just 0.8 yuan per million tokens and output costs of 3.2 yuan, slashing expenses to 40% of DeepSeek-V3’s levels. Meanwhile, the Wenxin X1 Turbo’s cost is just 25% of DeepSeek R1’s invocation price.

Baidu also announced China's first fully self-developed 30,000-card AI cluster, built using the third-generation Kunlun chip P800. Li said the system can simultaneously train multiple hundred-billion-parameter models like DeepSeek, while supporting fine-tuning for thousands of clients.

The push comes amid broader industry concerns. Li has previously criticized China’s "model race," arguing that too many domestic large models deliver little real-world value. In July 2024, he famously described open-source models as an "intellectual tax," claiming closed-source models like Baidu’s Ernie Bot or OpenAI’s ChatGPT offer superior power and lower costs.

While acknowledging the educational value of open-source models, Li stressed that for commercial applications, efficiency and cost-effectiveness favor closed ecosystems. He pointed out shortcomings in DeepSeek’s capabilities, including its inability to handle multimodal content and its high hallucination rate, warning of the risks such errors pose for businesses.

Capping off his remarks on Sunday, Li doubled down on his central belief: "Applications create the future, and developers shape the future."