Hunyuan-A13B - Tencent's open-source large language model based on the MoE architecture

What is Hunyuan-A13B?

Hunyuan-A13B is Tencent's latest open-source large language model, based on a hybrid expert (MoE) architecture with a total of 80 billion parameters and 13 billion activation parameters. It features a lightweight design and efficient inference capabilities, requiring only a single mid-to-low-end GPU card for deployment, significantly lowering the barrier to entry and making it suitable for individual developers and small and medium-sized enterprises. The model performs exceptionally well in mathematical, scientific, and logical reasoning tasks, providing step-by-step parsing for complex decimal comparisons. Users can choose between "fast thinking" or "slow thinking" modes to balance speed and accuracy. Hunyuan-A13B has released two open-source datasets, ArtifactsBench and C3-Bench, for code evaluation and agent-based scenario model evaluation, respectively, further promoting the development of the open-source ecosystem.

Main functions of Hunyuan-A13B

• Low-resource deploymentThe Hunyuan-A13B adopts a hybrid expert (MoE) architecture, which can be deployed with only one mid-to-low-end GPU card, significantly reducing inference latency and computational overhead, making it suitable for individual developers and small and medium-sized enterprises with limited resources.
• Mathematical and Logical ReasoningIt performs exceptionally well in mathematical reasoning tasks, such as accurately comparing the size of decimals and providing step-by-step solutions, and also demonstrates leading performance in scientific and logical reasoning tasks.
• Quick Thinking ModeSuitable for simple tasks, providing concise and efficient output, prioritizing speed and minimal computational overhead.
• Slow thinking modeSuitable for complex tasks involving deeper and more comprehensive reasoning steps, balancing efficiency and accuracy.
• Agent applicationsIt can call up tools to efficiently generate complex command responses such as travel guides and data file analysis, meeting diverse needs.
• Code evaluation and optimizationIt supports tasks such as code generation, debugging, and optimization through the open-source ArtifactsBench dataset, thereby improving programming efficiency.
• Intelligent Question AnsweringIt supports natural language processing tasks, such as text generation and question-answering systems, to provide users with accurate and helpful information.
• Open source supportThe model code has been open-sourced on GitHub, and users can freely download, modify, and use it, promoting community sharing and technological innovation.
• API AccessThe model API is now available on the Tencent Cloud website, making it easy for developers to quickly integrate and develop more application scenarios.

Technical Principles of Hunyuan-A13B

• Expert Hybrid (MoE) ArchitectureThe Hunyuan-A13B model employs the MoE architecture, with a total of 80 billion parameters and 13 billion activation parameters. By selectively activating relevant model components for each input, it significantly reduces inference latency and computational overhead, enabling deployment with only a single mid-to-low-end GPU card even under extreme conditions. Compared to dense models of similar scale, the MoE architecture offers advantages in inference speed and resource consumption.
• Pre-training and DatasetDuring the pre-training phase, Hunyuan-A13B utilized a high-quality corpus of 20 trillion network word units, covering multiple domains, significantly improving the model's generality and inference upper limit. The Tencent Hunyuan team refined the Scaling Law theoretical system of the MoE architecture, providing quantifiable engineering guidance for model design and further enhancing pre-training performance.
• Multi-stage training and optimizationIn the post-training phase, Hunyuan-A13B employs a multi-stage training approach to enhance the model's reasoning ability while also considering general capabilities such as creation, understanding, and agent functionality. The model supports a native 256K context window and performs exceptionally well in long text comprehension and generation tasks.

Project address of Hunyuan-A13B

• Github repositoryhttps://github.com/Tencent-Hunyuan/Hunyuan-A13B
• HuggingFace model libraryhttps://huggingface.co/tencent/Hunyuan-A13B-Instruct

Application scenarios of Hunyuan-A13B

• Agent applicationsThe Hunyuan-A13B can efficiently call tools to generate complex command responses, such as generating travel guides and analyzing data files, providing strong support for the development of intelligent agent applications.
• Mathematical and Logical ReasoningIn mathematical reasoning tasks, the Hunyuan-A13B performs exceptionally well, accurately performing decimal comparisons and providing step-by-step solutions.
• Long text comprehension and generationThe Hunyuan-A13B supports a 256K native context window and performs exceptionally well in long text comprehension and generation tasks, capable of handling complex text content.
• Code evaluation and generationUsing the open-source ArtifactsBench dataset, Hunyuan-A13B can be used for code generation, debugging, and optimization, covering multiple fields such as web development, data visualization, and interactive games.