Baidu Wenshin large model 4.5 series officially open source, synchronized open API services

Baidu's online shop, zhidao.baidu.comLarge model open source, as expected.

Just today, Baidu officially announcedERNIE4.5 series officially open source, also synchronized with the provision of API services.

This time, Baidu launched 10 open source models at once, covering everything from47BMixing specialists for parameters(MoE)Model to lightweight0.3BDensely modeled to cover a wide range of task requirements such as text and multimodality.

This open source not only weights and code completely open, but also synchronizes the provision ofAPI ServicesDevelopers can download and use it directly through Flying Paddle Star River Community, HuggingFace, and Baidu Intelligent Cloud Qianfan Platform.

△Wenshin big model 4.5 series open source models

Of interest is that the Wenshin Big Model 4.5 open source series follows theApache 2.0 protocol.

10 models synchronized open source

This time, Baidu launched 10 Wenxin Big Model 4.5 series open source models at once, and it has come up with sincerity in key dimensions such as the percentage of the number of independent self-research models, the number of model types, the richness of parameters, and the open source leniency and reliability.

Wenxin Big Model 4.5 open source series, also for the MoE architecture proposed an innovativeMultimodal heterogeneous model structure.

The structure is suitable for continuous pre-training paradigm from large language models to multimodal models, which significantly enhances the multimodal comprehension ability while maintaining or even improving the performance of the textual task, and its superior performance is mainly due to the key technical points of pre-training of multimodal hybrid expert models, efficient training inference framework, and modality-specific post-training.

In addition, the Wenshin Big Model 4.5 open source series both use theFlying Paddle Deep Learning FrameworkPerform efficient training, reasoning and deployment.

Model FLOPs utilization in pre-training of large language models(MFU)attainment47%.

△Wenxin 4.5 pre-trained models excel in mainstream benchmarks

The experimental results show that its series of models in multiple textual and multimodal benchmark tests achieveSOTAlevels, with particularly strong effects on instruction following, world knowledge memorization, visual comprehension, and multimodal reasoning tasks.

In terms of text modeling, the Wencent Big Model 4.5 open source series outperforms several mainstream benchmark reviews in theDeepSeek-V3,Qwen3and other models.

△Wenxin 4.5-300B-A47B model excels in mainstream benchmarks

In terms of multimodal models, the Wenshin Big Model 4.5 open-source series is based on powerful visual perception and rich visual common sense, realizing the unification of thinking and non-thinking, and outperforming the closed-source multimodal big models in the mainstream multimodal big model evaluation of visual common sense, multimodal reasoning, and visual perception.OpenAI o1.

In addition, on the lightweight model, the text center 4.5-21B-A3B-Base text model effect is comparable to that of the same-weightQwen3Considerably, the Wenshin 4.5-VL-28B-A3B multimodal model achieves SOTA among open-source models of the same magnitude, and can even compete with larger parameter modelsQwen2.5-VL-32BBye-bye.

△Multimodal post-training model achieves SOTA in multiple multimodal benchmark tests

Developer Benefits: Out-of-the-box Toolchain

It is understood that the Wenxin large model 4.5 open source series of weights in accordance with the Apache 2.0 protocol open source, support for academic research and industrial applications.

Additionally based on the fact that Flying Paddle offers an open source, industry-grade development kit, it can significantly reduce the post-training and deployment threshold of models due to its broad compatibility with a wide range of chips.

As one of the earliest enterprises to invest in AI research and development in China, Baidu has constructed a full-stack AI technology advantage in the four-layer layout of arithmetic, framework, model to application.

Among them, Flying Paddle, as China's first self-developed, feature-rich, open-source and open industry-grade deep learning platform, based on years of Flying Paddle open source technology and ecosystem accumulation, this Wenxin Big Model 4.5 open source series synchronously upgraded the release of Wenxin Big Model Development Kit ERNIEKit and Big Model Efficient Deployment Kit FastDeploy, to provide the Wenxin Big Model 4.5 series and developers with out-of-the-box tools and full-process support for Wuxin Big Model 4.5 series and developers.

In addition, it is worth mentioning that after the open source of the Wenxin Big Model 4.5 series, Baidu has also realized the framework layer and the model layer of the "two-layer open source".

(Source of article: Quantum Bits)