Baidu Just Leveled the Playing Field for AI Builders

3 minute read

Author

Midas AI Research
July 02, 2025

While most companies are still trying to catch up to GPT4, Baidu decided to go nuclear and open source the competition.

This week, the Chinese tech giant released Ernie 4.5, a suite of ten multimodal AI models with sizes ranging from a modest 0.3B to a massive 424B parameters. The most impressive part? They’re not just “comparable” to GPT4… they’ve outperformed Meta’s LLaMA 3 and Google’s Gemini in multiple benchmark tests. And unlike the closed walls of OpenAI, these models are now available for public use.

But Baidu didn’t stop at model release. They also rolled out ErnieKit and FastDeploy, a pair of developer tools designed to fine tune, integrate, and deploy the models into real world business applications without any reliance on OpenAI’s infrastructure. The kicker? It all runs on Baidu’s proprietary 30,000 chip cluster… built in house, powered by their own silicon.

This isn’t just a tech flex. It’s a strategic move aimed squarely at startups, enterprise teams, and global developers who have been boxed in by expensive API calls and limited flexibility. Baidu is saying, here’s the code, here are the tools, go build something better and faster.

For startups and builders, this could be a turning point. The smaller models in the Ernie suite… especially the 21B parameter version… are already beating GPT3.5 in tasks like customer support, summarization, and product recommendation. And the fact that they're open-source means companies can fine tune and deploy locally, without waiting on OpenAI pricing changes or terms of service.

Whether or not Baidu wins the global AI arms race, they’ve just made a move that forces everyone… especially businesses reliant on closed models to reconsider what’s possible.

Sources: