ERNIE-4.5-21B-A3B-Thinking - Baidu open source reasoning thinking model

What is ERNIE-4.5-21B-A3B-Thinking?

ERNIE-4.5-21B-A3B-Thinking is Baidu open source large-scale language model focused on reasoning tasks. Using the Mixed Expert (MoE) architecture, the total number of references up to 21 billion, each token activates 3 billion parameters, supports 128K long context window, suitable for complex reasoning tasks. The model builds the core backbone of the language through textual pre-training, and in the post-training phase of inference enhancement, it utilizes techniques such as Supervised Fine Tuning (SFT) and Progressive Reinforcement Learning (PRL), which significantly improves the ability of logical inference, mathematical computation, and scientific question answering. It supports efficient tool invocation and can be integrated with vLLM, Transformers 4.54+, and FastDeploy for scenarios such as program synthesis, symbolic reasoning, and multi-intelligence body workflow.

Functional Features of ERNIE-4.5-21B-A3B-Thinking

• Hybrid Expert Architecture: Adopting MoE architecture, the total number of references is 21B, and each token activates 3B parameters, balancing performance and efficiency.
• Long Context Processing: Supports 128K context windows, suitable for long text reasoning tasks such as complex document analysis.
• Reasoning Strengthening: Logical reasoning, mathematical calculations, and scientific problem solving are significantly improved with reasoning enhancement training.
• Tool call support: Supports structured tools and function calls, can be integrated with vLLM, Transformers 4.54+, etc. to expand application scenarios.
• Open source and easy to deploy: Open source under the Apache-2.0 license, available on platforms such as Hugging Face for research and commercial deployment.

Core Benefits of ERNIE-4.5-21B-A3B-Thinking

• Efficient reasoning skills: excels in complex tasks such as logical reasoning, mathematical calculations, and scientific problem solving, giving quick and accurate answers.
• long contextual understanding: Supports 128K long context windows, capable of handling long text messages for complex tasks that require long contextual understanding.
• Hybrid Expert Architecture: Adopting MoE architecture with 21B total references and 3B parameters activated by each token, it takes into account both performance and efficiency, and has a high utilization of computational resources.
• open source and easy to use: Open source under the Apache-2.0 license, available on platforms such as Hugging Face, for developers' research and commercial deployment.
• Tool call support: Supports structured tool and function calls, can be integrated with vLLM, Transformers 4.54+, etc. to expand application scenarios.

What is the official website for ERNIE-4.5-21B-A3B-Thinking?

• HuggingFace Model Library:: https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking

People for whom ERNIE-4.5-21B-A3B-Thinking is suitable

• (scientific) researcher: ERNIE-4.5-21B-A3B-Thinking's powerful reasoning ability and long contextual comprehension can assist researchers in complex logical reasoning, scientific problem exploration, and academic research, helping them to process and analyze large amounts of literature more efficiently and accelerate the research process.
• developers: The open-source features and support for tool calls allow developers to easily integrate them into various applications for the development of intelligent code generation, automated programming assistance and other functions to improve development efficiency, and to add intelligent analysis and other features to software products by means of the model's reasoning capabilities.
• educator: Models can be used to assist teaching and learning, such as generating practice problems, answering students' questions, and providing learning suggestions, especially in teaching subjects that require reasoning and logical thinking, such as math and science, to provide students with personalized learning support.
• business user: In the business field, enterprises can process complex business data, market analysis reports, etc. based on reasoning and analyzing capabilities to assist decision-making, for example, providing valuable insights in financial risk assessment, market trend forecasting, etc., to enhance the competitiveness of enterprises.
• technology enthusiast: For technology enthusiasts interested in artificial intelligence and natural language processing, ERNIE-4.5-21B-A3B-Thinking is a good object of study and experimentation, and they can expand their technical horizons by studying and using the model to gain a deeper understanding of the workings and application scenarios of large-scale language models.