Baidu drops ERNIE 4.5 model family with up to 424B parameters

Baidu has landed its ERNIE 4.5 family on Hugging Face, open-sourcing a full lineup that ranges from lightweight 0.3B edge models all the way up to a 424B multimodal heavyweight. If you need a capable vision-language foundation with pre-built quantizations and dedicated reasoning variants, this collection provides an extensive toolkit for local fine-tuning and deployment.

From edge scale to a 424B flagship

The released ERNIE 4.5 collection is split across text generation and vision-language multimodal architectures.

The flagship multimodal checkpoint is ERNIE-4.5-VL-424B-A47B, packing 424B total parameters with roughly 47B activated parameters. On the pure text side, the largest available variant is the ERNIE-4.5-300B-A47B model.

For more accessible workloads, Baidu packaged mid-sized options including the 28B multimodal ERNIE-4.5-VL-28B-A3B and the 21B text model ERNIE-4.5-21B-A3B. At the extreme low end, a 0.3B parameter model is also included for ultra-lightweight edge setups.

PyTorch weights and Thinking variants

Instead of restricting checkpoints to its proprietary PaddlePaddle framework, Baidu released weights in both PaddlePaddle and standard PyTorch formats across all tiers, eliminating friction for existing pipelines.

For complex reasoning tasks, the release features specialized Thinking editions: ERNIE-4.5-21B-A3B-Thinking and the multimodal ERNIE-4.5-VL-28B-A3B-Thinking, both tuned for extended step-by-step logic generation.

To help developers run massive models without datacenter-grade clusters, Baidu also uploaded multiple compressed variants for the 300B model, including FP8, W4A8C8 tensor-parallel builds, and extreme 2-bit quantized weights.

Where to run it

All checkpoints, configurations, and tokenizer files are live under the official baidu organization on Hugging Face inside the ernie-45 collection.

Developers can pull either pre-trained Base checkpoints or instruction-tuned chat checkpoints directly via the Hugging Face hub, making them straightforward to integrate with standard inference engines.

Why it matters

Open multimodal models crossing the 400B parameter mark remain rare in the open-source landscape. The 21B and 28B Thinking versions hit an ideal sweet spot for teams with limited hardware that still demand serious reasoning capabilities.

We think releasing native PyTorch checkpoints alongside aggressive 2-bit quantization shows Baidu is getting pragmatic about developer adoption. If your team is evaluating high-capacity open-weight models for enterprise multimodal tasks, this collection is well worth spinning up for a benchmark run.