Recently, the ERNIE large model family under Baidu has undergone a major upgrade - ERNIE-4.5-21B-A3B-Thinking has been officially open-sourced and quickly topped the text generation model ranking on the Hugging Face platform, while also ranking third in the overall model list. This lightweight Mixture-of-Experts (MoE) model has attracted widespread attention in the industry with its outstanding reasoning capabilities and parameter efficiency, marking another milestone in China's AI open-source ecosystem.

Core Specifications and Innovative Design

ERNIE-4.5-21B-A3B-Thinking uses an advanced MoE architecture, with a total parameter scale of 21B, but only 3B parameters are activated per token. This sparse activation mechanism significantly reduces computational costs while maintaining high performance output. The model supports a 128K long context window, making it particularly suitable for processing complex long-text tasks such as logical reasoning, mathematical problem-solving, and academic analysis.

image.png

Differing from mainstream models that rely on the PyTorch framework, the ERNIE-4.5 series is trained and optimized based on Baidu's self-developed PaddlePaddle deep learning framework. This independent framework design not only enhances the compatibility of the model in multimodal tasks, but also ensures efficient hardware adaptation. Currently, only Baidu and Google use self-developed frameworks to train large models worldwide, highlighting their technical autonomy and innovation depth.

Performance: Efficient Reasoning Challenging Industry Giants

According to the latest benchmark tests, the model performs excellently in tasks such as logical reasoning, mathematics, science, coding, and text generation, even approaching or surpassing models at the level of Gemini 2.5 Pro and GPT-5 in some metrics. Despite a total parameter count of only 21B (approximately 70% of Qwen3-30B), its scores on mathematical reasoning benchmarks like BBH and CMATH have already exceeded those of competing models, demonstrating extremely high parameter efficiency.

In addition, the model includes an efficient tool calling function, supporting structured function calls and external API integration, suitable for program synthesis, symbolic reasoning, and multi-agent workflow scenarios. In terms of long-context understanding, after specialized training, it can stably process massive information to generate academic-level synthetic content, significantly reducing hallucination issues. The model also supports bilingual (Chinese and English) optimization, suitable for global developers and enterprise applications.

Feedback from the open-source community shows that the model's download volume and trend index have surged on Hugging Face, becoming a popular choice in the field of text generation. Developers can easily integrate it using tools such as vLLM, Transformers 4.54+, and FastDeploy, achieving local deployment or cloud-based inference.

Open Source Significance: Promoting AI Democratization and Ecosystem Development

ERNIE-4.5-21B-A3B-Thinking is released under the Apache 2.0 license, supporting commercial use, further lowering the barrier to AI technology. Following the open-sourcing of other 10 models in the ERNIE 4.5 family at the end of June, this release strengthens Baidu's leadership in the open-source AI field. Currently, many top models on the Hugging Face platform are Chinese open-source achievements, reflecting China's global competitiveness in MoE architecture and reasoning optimization.

As the latest iteration of the ERNIE large model, this model not only improves performance in instruction following and knowledge-intensive tasks, but also enhances its "thinking" mode through multi-round reinforcement learning. In visual-language tasks, its VL variant also performs well, narrowing the gap with OpenAI-o1 on benchmarks such as MathVista and MMMU.

Industry Impact and Future Prospects

The release of this model proves that it is possible to achieve deep reasoning without trillions of dense parameters. It provides high-performance options for developers with limited resources, driving the transformation of AI from laboratories to practical applications. In the future, as the PaddlePaddle framework continues to expand its ecosystem, the ERNIE series is expected to play a greater role in agent products and multimodal applications, avoiding risks brought by monopolies from single vendors.