On June 30, 2025, Baidu made its Ernie 4.5 family of large-scale AI models freely available under the Apache 2.0 license. This landmark release includes ten model variants built on a Mixture-of-Experts (MoE) architecture and accompanying multi-hardware development toolkits, empowering researchers and developers worldwide to innovate without licensing barriers.
Ernie 4.5 Model Family at a Glance
- Ten Variants: Four multimodal vision-language models, eight MoE models, and two reasoning-focused models.
- Training Status: Five variants are post-trained, while the remainder remain in their original pre-trained state.
- Parameter Counts: MoE models span from 47 billion total parameters (with 3 billion active at runtime) up to a 424 billion-parameter behemoth.
- Framework: All models leverage Baidu’s in-house PaddlePaddle deep learning stack for both training and inference.
Architectural Innovations
Baidu’s Ernie 4.5 series pioneers a heterogeneous MoE structure that balances shared parameters across modalities with expert-specific weights. Key techniques include:
- Modality-Isolated Routing: Ensures text and vision experts don’t conflict during joint training.
- Router Orthogonal Loss: Promotes diversity among expert modules.
- Multimodal Token-Balanced Loss: Keeps training data distribution in check across text and image inputs.
- Scaling-Efficient Infrastructure: Combines intra-node expert parallelism, FP8 mixed-precision training, memory-efficient pipeline scheduling, and fine-grained recomputation to maximize throughput.
These choices yield state-of-the-art performance on text understanding, visual reasoning, and cross-modal tasks.
Benchmark Performance
Internal tests highlight the Ernie 4.5 models’ competitive edge:
- Ernie-4.5-300B-A47B-Base outperformed DeepSeek-V3-671B-A37B-Base on 22 of 28 standard benchmarks.
- Ernie-4.5-21B-A3B-Base surpassed Qwen3-30B-A3B-Base in mathematics and reasoning tasks despite having 30% fewer parameters.
Such metrics underline the efficiency of MoE routing in delivering high performance with leaner compute budgets.
Open-Source Licensing & Access
All Ernie 4.5 weights and toolkits are published under the permissive Apache 2.0 license, enabling both academic study and commercial deployment. Models and resources can be downloaded from:
- Baidu’s GitHub repository
- Hugging Face model hub
This unrestricted access marks a pivotal move in China’s AI landscape toward greater openness and community collaboration.
Multi-Hardware Development Toolkits
Alongside model weights, Baidu released ErnieKit, a modular toolkit supporting:
- Pre-Training & Supervised Fine-Tuning (SFT)
- Low-Rank Adaptation (LoRA) and other parameter-efficient customization
- Resource-Efficient Inference Pipelines on CPUs, GPUs, and specialized accelerators
Toolkits are designed for multi-hardware compatibility, with phased rollout ensuring developers can gradually scale from local prototypes to cloud-based production environments.
Implications for Developers
By open-sourcing Ernie 4.5 and its toolchains, Baidu:
- Lowers Entry Barriers for AI research and productization
- Fosters Ecosystem Growth, inviting contributions to model improvements and benchmark suites
- Encourages Hardware Diversity, from edge devices to data-center clusters
Developers in academia and industry can now integrate cutting-edge multimodal AI into applications ranging from conversational agents to computer-vision systems without licensing friction.
Future Directions
Baidu’s open-source push sets the stage for:
- Community-Driven Extensions: Custom module integrations, benchmark contributions, and optimized deployment scripts.
- Industrial Adoption: Seamless integration into enterprise workflows across finance, healthcare, and beyond.
- Ecosystem Lock-In: As more teams adopt PaddlePaddle and ErnieKit, competitor frameworks may face steeper switching costs.
This strategic release not only cements Baidu’s AI leadership in China but also enriches the global open-source AI community.
