Baidu Inc. has intensified the global AI arms race with the release of 23 new foundation models—ranging from 0.3 billion to 424 billion parameters—on Hugging Face, the world’s largest platform for open machine learning models. The move, revealed in a blog post late Sunday, is the clearest signal yet that China’s search giant is aiming to play in the same league as OpenAI, Meta, and Google—on their turf, and by their rules.

The suite, branded under the ERNIE 4.5 moniker, includes a mix of dense, distilled, and mixture-of-experts (MoE) architectures, all open-sourced under the permissive Apache 2.0 license. Most notable is the ERNIE 4.5-VL-424B-A47B, a sparse multimodal model that activates 47 billion parameters per token—making it among the largest publicly available AI models of its kind.

Baidu’s release strategy leans heavily on breadth and optionality. The 23-model drop includes:

• A 0.3B parameter dense model tuned for low-latency applications;

• Mid-weight MoE variants (21B–128B total) for instruction following and chat;

• The 424B multimodal model, built for text-and-vision tasks, with a 128K token context window;

• Tool-use and function-calling variants tuned for real-world deployment.

All models are trained on Baidu’s PaddlePaddle deep learning framework and come with tooling for vLLM, FastDeploy, and other industry-standard inference backends.

The architecture of the flagship models is particularly revealing. ERNIE’s MoE design enables the activation of a subset of experts per input—typically 2 out of 64—drastically reducing computational overhead while preserving parameter scale. Combined with FP8 mixed precision, 4-bit quantization, and hierarchical load balancing, the models are built not just for power but for efficiency at scale.

The release lands amid rising scrutiny over the closed nature of U.S.-based frontier models. By publishing on Hugging Face and aligning with open-weight culture, Baidu is positioning itself as a viable alternative to Meta’s LLaMA 3 and Mistral’s Mixtral suite.

This pivot toward openness follows earlier signals. In March, Baidu promised to open-source ERNIE by June 2025 after backlash over restricted licensing. With the 4.5 series now live on Hugging Face, Baidu appears to be delivering early—and with a full portfolio, not just a single headline-grabber.

The models are also being offered free of charge for individual users, with enterprise pricing reportedly 90% cheaper than GPT-4 Turbo, according to Baidu executives. That move will likely pressure global incumbents to further rethink pricing and distribution strategies.

Baidu’s decision to lean into open-source and multimodal research is not just a technical milestone—it’s a geopolitical one. As U.S. firms face increasing regulatory scrutiny and chip constraints, China’s largest AI players are doubling down on scale, training infrastructure, and access.

By releasing models with SOTA-level benchmark claims and versatile deployment tooling, Baidu is attempting to seed its ecosystem beyond China’s borders. And by putting them on Hugging Face, it’s targeting the attention of Western developers, researchers, and VCs.

Baidu’s ERNIE 4.5 release is a bold gambit to compete not just on performance, but on openness and developer mindshare. With 23 models, including one of the largest publicly available MoE architectures, Baidu has made clear it intends to be more than China’s answer to OpenAI—it wants to be the next foundational layer of AI itself.

TLDR
Baidu has released 23 ERNIE 4.5 models—ranging from 0.3B to 424B parameters—on Hugging Face, signaling a new phase in China’s push for AI supremacy. With open licenses, multimodal capabilities, and competitive inference costs, Baidu is now vying for global developer and enterprise adoption in the open-weight model wars.