ERNIE 4.5 and ERNIEKit: Baidu's open-source MoE and VLM training stack on PaddlePaddle

What ERNIE 4.5 and ERNIEKit Actually Solve

ERNIE 4.5 is a family of multimodal models released by Baidu under Apache-2.0, and ERNIEKit is the accompanying development toolkit. The core problem it addresses is training and fine-tuning large MoE models that span text, image, and video inputs without forcing you to assemble a custom distributed training stack. The model family includes 10 variants, from a 0.3B dense model to a 424B total parameter MoE with 47B active parameters. The toolkit targets engineers who need SFT, LoRA, and function calling support on these specific architectures, not generic LLM experimentation. The README positions ERNIEKit as an industrial-grade toolkit, which suggests it is meant for production workloads rather than research prototypes. If you are working outside the PaddlePaddle ecosystem, the value proposition drops sharply, because the models and toolkit are tightly coupled to that framework.

Heterogeneous MoE Architecture and Context Window

The ERNIE 4.5 models use a Mixture-of-Experts design with a heterogeneous modality structure. The README explains that this allows parameter sharing across modalities while keeping dedicated parameters for each modality. That is a specific design choice aimed at improving multimodal understanding without hurting text-only performance. All models listed have a 128K context window, which is uniform across LLMs, VLMs, and dense models. The largest model, ERNIE-4.5-VL-424B-A47B, has 424B total parameters and 47B active, meaning inference requires fewer resources than a dense model of similar total size. The 28B-A3B variant is notable because ERNIEKit supports SFT and function calling training for it, as announced in the v1.5 release. The architecture is not novel in the broad sense, but the explicit heterogeneous sharing is a differentiator compared to models that treat modalities as separate towers.

ERNIEKit Releases and What Each Version Adds

ERNIEKit has progressed quickly through versions. v1.0 arrived in June 2025, and by November 2025 it reached v1.5. Each release adds specific training capabilities. v1.1 added SFT and LoRA for the ERNIE-4.5-VL series. v1.2 added WebUI support for training and conversation with the 28B and 424B VL models, plus a query-response format for training data and iluvatar GPU support. v1.3 added SFT and function calling for the ERNIE-4.5-21B-A3B-Thinking model. v1.4 introduced SFT for the PaddleOCR-VL-0.9B model and a padding-free dataflow strategy that packs data within a batch to avoid padding, reducing GPU memory and speeding up training. v1.5 added SFT and function calling for ERNIE-4.5-VL-28B-A3B-Thinking. The cadence suggests active development, but it also implies that APIs and configuration formats may shift between minor releases. Teams should pin to a specific version rather than tracking the default branch.

Getting Started: Commands and Configuration

The README points to a training document at docs/erniekit.md for setup and usage. It also references a cookbook directory for examples. The repository does not show a single pip install command in the README, so you must consult the training document for exact installation steps. The documentation mentions multi-hardware compatibility, including iluvatar GPUs from v1.2 onward. For data, the v1.2 release introduced a query-response format for VL training, which likely means your dataset must follow a specific JSON structure. The padding-free strategy in v1.4 is a dataflow optimization that you enable during training configuration. To get a concrete start, clone the repository, check the docs/erniekit.md file, and look at the cookbook for example scripts. The absence of a quickstart in the README is a friction point; you cannot just run a single command to start training.

Limitations and Failure Modes

The most obvious limitation is the tight coupling to PaddlePaddle. If your existing stack is PyTorch-based, adopting ERNIE means maintaining a second deep learning framework, which is a significant operational cost. The README claims high MFU (47% for the largest pre-training run), but that is a pre-training metric, not something you will reproduce during fine-tuning. Another limitation is the youth of ERNIEKit. The first release was in June 2025, so the toolkit has less than a year of maturity. Bug fixes in v1.2 and v1.3, such as AutoParallel checkpoint and pp+recompute+moe issues, indicate that distributed training edge cases are still being ironed out. The 128K context window is fixed across all models, so if you need a longer context, this family is not the right fit. Also, the VL models output only text, not images or video, which constrains what you can build for multimodal generation.

Alternatives and How They Differ

The closest alternative is Hugging Face Transformers with a PyTorch backend, which supports many open MoE models like Mixtral or Qwen. The difference is that Hugging Face provides a framework-agnostic API, while ERNIEKit is built specifically for PaddlePaddle and the ERNIE architectures. If you want to fine-tune a model without learning a new toolkit, Hugging Face's Trainer or SFTTrainer offers a more familiar path. Another alternative is vLLM for inference, which is popular for serving LLMs, but it does not cover training. ERNIEKit includes training and inference workflows, so it is more of an all-in-one stack. The trade-off is that you trade ecosystem breadth for depth in a single vendor's stack. For teams already using PaddlePaddle for other models, ERNIEKit is a natural extension; for everyone else, the switching cost is high.

Maintenance, Licensing, and Upgrade Considerations

The repository is licensed under Apache-2.0, which is permissive and allows commercial use, modification, and redistribution without copyleft obligations. The models themselves are also released under Apache-2.0, as stated in the README. The default branch is release/v1.5, and the last push was July 2026, indicating active maintenance. However, the rapid release cycle means you should expect to upgrade frequently to get bug fixes and new features. Each minor version has introduced new model support, so if you adopt an older version, you will miss out on the latest VL training capabilities. The documentation is spread across docs/erniekit.md and a cookbook directory, which is adequate but not as polished as some commercial offerings. Before upgrading, review the release notes for breaking changes, especially around AutoParallel and data format, as those areas have seen fixes.

Editorial conclusion

Adopt ERNIE 4.5 and ERNIEKit if you are already invested in PaddlePaddle, need a MoE model with a 128K context window, or want to fine-tune a vision-language model with SFT and function calling support. Do not adopt it if you require a pure PyTorch workflow or need extensive community plugins. Before committing, verify that your GPU drivers and PaddlePaddle version match the ERNIEKit release, and check the cookbook for the exact data format expected by the training scripts. The repository is actively maintained, but the toolkit is young; plan for breaking changes between v1.x releases.

Official sources

1. License: Apache-2.0
2. PaddlePaddle/ERNIE on GitHub
3. Project website
4. README
5. Releases