One day a model is a paid cloud API. The next it is free to download, fine-tune, and ship inside your own product. That is the line Baidu crossed with ERNIE 4.5 (文心4.5).
On June 30, 2025, Baidu open-sourced the entire ERNIE 4.5 family, releasing pretrained weights and inference code for 10 models under the permissive Apache 2.0 license. The move was confirmed by Xinhua and by Baidu's own technical blog.
What exactly got opened
Baidu did not open a single model — it opened a ladder:
- 10 models in total.
- Mixture-of-Experts (MoE) variants with 47B and 3B active parameters, the largest with 424B total parameters.
- A lightweight 0.3B dense model for edge and mobile use.
- Both text and vision-language (VL) versions, the latter supporting "thinking" and "non-thinking" modes.
The weights are available on the PaddlePaddle (飞桨) community and on Hugging Face, and the open-source API service runs on Baidu's Qianfan (千帆) platform. Baidu says the family uses a multimodal heterogeneous MoE structure that shares parameters across text and vision while keeping modality-specific experts — a design meant to lift multimodal understanding without hurting text performance.
The strategy behind the release
Baidu previewed the ERNIE 4.5 plan in February 2025 and first showed the models in spring 2025 before opening the weights at the end of June. The company framed it as "dual-layer open source": it opened not just the model but also the tooling — ERNIEKit for fine-tuning and alignment, and FastDeploy for efficient serving — plus broad hardware compatibility through PaddlePaddle.
The commercial logic is familiar in the open-model era: give developers the weights for free, then monetize through cloud hosting, enterprise support, and ecosystem lock-in. Baidu reports a Model FLOPs Utilization (MFU) of 47% in the largest language-model pretraining, a metric it uses to signal training efficiency.
What Baidu claims on benchmarks
Baidu's own reporting, echoed by Xinhua, asserts the models beat or match strong peers:
- The 300B-class base model surpasses DeepSeek-V3 on a majority of 28 benchmarks.
- The lightweight 21B-class model outperforms Qwen3-30B on several math and reasoning tests despite fewer parameters.
- The vision-language model narrows or closes the gap to OpenAI's o1 on hard multimodal reasoning sets such as MathVista and MMMU.
These are vendor benchmark claims. They are useful signals of relative capability, not independent verdicts, and benchmark leadership in this segment moves month to month.
The bigger picture: China's open-weight wave
ERNIE 4.5 joins a crowded open-release field from Chinese labs — Alibaba's Qwen, DeepSeek, Zhipu's GLM, Moonshot's Kimi and others have all published weights. For Baidu, a company that spent two decades building a full AI stack (chips, framework, model, application), opening ERNIE is partly defensive: it keeps developers inside the PaddlePaddle ecosystem rather than letting them drift to a rival's open model.
What changes for developers on day one
The practical effect of Apache 2.0 is that a team can download the weights, run them on its own GPUs, and ship a product without per-call API fees. Baidu's FastDeploy and ERNIEKit are meant to lower the serving and fine-tuning barrier, and the company highlights multi-hardware compatibility beyond its own cloud. The irony for developers is the opposite of lock-in: with so many open Chinese models now available, the hard part becomes evaluation and operations — not access.
The part that is easy to miss: the framework play
Opening a model is also a statement about a framework. ERNIE 4.5's weights were trained, aligned and served through Baidu's own PaddlePaddle (飞桨) stack, and the release is designed so that fine-tuning a downloaded checkpoint naturally pulls a team deeper into that ecosystem. In the US, open weights from Meta entrenched PyTorch; in China, the same logic now runs through PaddlePaddle against PyTorch-based alternatives. Every team that adopts ERNIEKit for alignment tooling is a data point in that quiet infrastructure contest — which is one reason the release included the tooling, not just the tensors.
Honest limitations
- Performance comparisons (vs DeepSeek-V3, Qwen3, OpenAI o1) are Baidu's own benchmark results reported by Xinhua and Baidu; we did not re-run them.
- The "first / most open" framing in coverage is Baidu's; other Chinese labs have also released large open-weight families.
- Apache 2.0 permits commercial use, but real deployment still depends on hardware, serving cost, and compliance — none of which a license alone solves.
- We relied on Xinhua and Baidu's official blog for the model roster and license; we did not inspect the model cards directly.
What readers can do now
- If you build with Chinese open models, benchmark ERNIE 4.5-21B-A3B against Qwen3-30B on your own tasks before committing — published scores rarely match your domain.
- Check the PaddlePaddle and Hugging Face model cards for license terms, context length, and quantization support before production use.
- Watch whether Baidu follows with open smaller "edge" variants; the 0.3B dense model hints at an on-device push worth tracking.