Just when you thought DeepSeek was running away with the open-source crown, Baidu drops an absolute bombshell.
I’ve been watching the AI leaderboards closely, and this is huge. The Chinese tech giant just open-sourced its ERNIE 4.5 model family, and it’s a serious contender.
This isn’t just another model release; it’s a major power move. Baidu’s CEO was skeptical about open-sourcing just a year ago, but now they’re all in.
Here’s the lowdown
- A New Champion? Their biggest model, a 424-billion-parameter multimodal system, beats DeepSeek V3 on 22 out of 28 benchmarks. The craziest part? It’s half the size. That’s insane efficiency.
- Smart Design: It uses a Heterogeneous training architecture. In simple terms, the text and vision parts of the model work together to boost each other, not compete for resources. This is how they get so much power from a smaller model.
- Free for All: They released 10 different versions on Hugging Face, from tiny to massive, all under the friendly Apache 2.0 license. This is Baidu’s first big open-source play, and they came to win.
This move supercharges the AI rivalry in China. You have Baidu, DeepSeek, Alibaba, and others all pushing the pace, which means we get more powerful and awesome open-source tools to play with.
What a time to be building!