Chinese tech giant Baidu officially released an open source family of Ernie 4.5 language models this week, showing a major change in the company's AI strategy, potentially disrupting the global AI landscape.
This release includes ten different variations of the multimodal model family, ranging from a lightweight 0.3 billion parameter model to a heavyweight system with 4240 billion parameters. Currently, all models are available under the Apache 2.0 license and are free to use, modify and distribute.
According to Baidu's technical report, Ernie 4.5 features three important innovations:
First, it features a multimodal, non-uniform mixture pre-training architecture that improves the performance of text understanding, image understanding, and cross-modal inference.
Second, there is a scaling-efficient infrastructure that allows for high-throughput training and inference.
Third, there is modality-specific post-training training that optimizes the model for a particular use case.
Baidu cited the benchmark test. This states that the 300B Ernie 4.5 model outperforms the Deepseek V3 model despite being half its size. The company reports high performance, memory of world knowledge, visual understanding, and multimodal reasoning in the following teaching.
This release represents a dramatic reversal for Baidu. Baidu previously said CEO Robin Li is more powerful than open source alternatives, such as Ernie.
The change, following Deepseek's January open source release, has gained significant traction in the AI community, and accelerated the shift towards an open source model for Chinese tech companies.
“Baidu can prove that LLM Ernie can prove an earthquake more than Deepseek's January bomb,” said Sewa Rejal, chief commercial officer of AI infrastructure company Netmind.
“While Deepseek has been a proof of the relatively unknown startup concept, Baidu brings institutional weight, capital firepower, and, crucially, distribution channels to ensure widespread adoption. Despite US chip restrictions, China is able to release cheap, high-performance AI models.
This release includes a development toolkit built on Baidu's PaddlePaddle Deep Learning Framework (Andiekit for fine-tuning and alignment, and FastDeploy for efficient model deployment). The company has also made the Pytorch compatible version available to accommodate developers in its ecosystem.
Baidu joins other Chinese tech giants that embrace open source.
Alibaba's Qwen model has become the world's most popular open source model among developers, but Huawei Technologies opened sourced two of the Pangu AI models on the same day as Baidu's announcement.
This trend is putting pressure on Western AI companies, including Openai and humanity, who are primarily maintaining their own closed models. Openai CEO Sam Altman confirmed earlier this year that the company needs to “know a different open source strategy” and committed to releasing an open source model, but a recent update shows that this release was delayed.
The Ernie 4.5 model family includes special variants for visual language understanding, with both “thinking” and “non-thinking” modes. According to Baidu, the thinking mode increases reasoning ability while maintaining powerful perceptual abilities, resulting in consistent results with a variety of multimodal evaluation benchmarks such as Mathvista, Mmmu, and VisualPuzzle.
For developers, this model offers industrial grade functionality through Erniekit. It offers model training and compression capabilities including pre-training, monitored fine-tuning, low-rank adaptation, direct-priority optimization, and a variety of quantization techniques. The FastDeploy Toolkit offers multi-hardware deployments with single code compatibility and API compatibility with both VLLM and OpenaI protocols.
