Baidu launched the Ernie 4.5 collection of synthetic intelligence (AI) fashions in open-source on Monday. The Chinese language tech large had beforehand stated that it will make its proprietary giant language fashions (LLMs) accessible to the open group on July 31. It has now launched 10 totally different variants of the collection, with every of the fashions constructed on Combination-of-Consultants (MoE) structure. Alongside the fashions, the corporate has additionally launched multi-hardware growth toolkits for Ernie 4.5 in open supply.

Baidu Releases 10 Variants of Ernie 4.5 AI Fashions in Open Supply

In a post on X (previously often called Twitter), the Chinese language tech large introduced the discharge of the ten open-source Ernie 4.5 AI fashions. 4 of them are multimodal vision-language fashions, eight are MoE fashions, and two are considering or reasoning fashions. Moreover, the listing additionally consists of 5 post-trained fashions, whereas others are pre-trained. These fashions can now be downloaded from both the corporate’s Hugging Face listing or from its GitHub listing.

In a blog post, Baidu stated that the MoE fashions function a complete of 47 billion parameters, with three billion of them being energetic at a time. The most important fashions among the many 10 variant options 424 billion parameters. All of them are educated utilizing the PaddlePaddle deep studying framework.

Based mostly on inner testing, the corporate claimed the Ernie-4.5-300B-A47B-Base mannequin surpasses DeepSeek-V3-671B-A37B-Base on 22 out of 28 benchmarks. Equally, it claimed that the Ernie-4.5-21B-A3B-Base outperforms Qwen3-30B-A3B-Base on a number of arithmetic and reasoning benchmarks regardless of having 30 % fewer parameters.

Baidu additionally revealed its coaching strategies on the mannequin pages. The corporate used a heterogeneous MoE construction within the pre-training course of and scaled the fashions utilizing methods equivalent to intra-node skilled parallelism, memory-efficient pipeline scheduling, FP8 mixed-precision coaching, and a fine-grained recomputation methodology.

Aside from the fashions, Baidu has additionally made ErnieKit accessible to the open group. It’s a growth toolkit for the Ernie 4.5 collection fashions. With this, builders can carry out pre-training, supervised fine-tuning (SFT), Low-Rank Adaptation (LoRA), and different customisation methods. Notably, all of the fashions can be found below the permissive Apache 2.0 licence, which permits for each tutorial and industrial utilization.