Author: Zhou Yuan / Wall Street News
On June 27, Tencent Hunyuan announced the open-source of its first hybrid inference MoE (Mixture of Experts) model Hunyuan-A13B, along with the release of two new datasets, ArtifactsBench and C3 - Bench, providing new technical resources and evaluation tools for the development of large models.
The Hunyuan-A13B model has a total of 80 billion parameters (80B) and 13 billion active parameters (13B), which gives it certain advantages in inference efficiency.
Compared to open-source models with the same architecture, using the common Transformer architecture model as an example, Hunyuan-A13B shows significant improvements in inference speed and relatively lower computational resource consumption when handling tasks of the same scale.
As the first open-source 13B-level MoE hybrid inference model, it has demonstrated certain general capabilities in multiple authoritative industry data tests, particularly excelling in agent tool invocation and long text processing, which gives it differentiated competitive advantages in practical application scenarios.
Tencent Hunyuan enhances the tool invocation capability of Hunyuan-A13B by constructing a multi-agent data synthesis framework.
This framework integrates various environments such as MCP (Large Model Context Protocol), sandbox, and large language model simulation, and employs reinforcement learning mechanisms to allow agents to learn in different environments.
In a travel scenario, when a user inputs the command &#34;plan a trip from Chengdu to Sichuan,&#34; the model can invoke map search tools to obtain route information, call hotel booking platforms to filter suitable accommodations, and use weather query tools to understand the weather during the trip, ultimately outputting a detailed itinerary that includes daily arrangements, transportation methods, accommodation recommendations, and attraction introductions.
In data analysis tasks, when faced with sales data from an e-commerce platform, the model can invoke Python coding tools for data cleaning and statistical analysis, generating an Excel sales analysis report that includes charts, meeting the complex task requirements of users in different scenarios.
Compared to some models that only have single tool invocation capabilities, Hunyuan-A13B's multi-tool collaborative invocation ability can better solve practical problems.
Facing the challenges of long text processing in large models, Hunyuan-A13B supports a 256K native context window.
In the academic field, when processing academic papers of tens of thousands of words, the model can accurately extract the core viewpoints of the paper, outline research methods and experimental results; in the legal industry, it can quickly summarize legal points and relate relevant legal provisions when analyzing complex legal texts and case files; in the business field, it can accurately extract key data and market trend information when interpreting lengthy business reports.
In practical tests, compared to some models with smaller context windows that easily miss information when processing long texts, Hunyuan-A13B alleviates the issues of context loss and information dependency in long text inference to a certain extent, providing more reliable technical support for applications in related fields
Hunyuan-A13B is open-source and relatively friendly to developers.
Individual developers can deploy it using a mid-range GPU card, such as the NVIDIA GeForce GTX series, under certain conditions.
Currently, the model has been integrated into mainstream open-source inference framework ecosystems and supports various quantization formats, including INT4 and INT8. With the same input-output scale, its overall throughput capacity reaches twice that of leading open-source models.
Developers can access the model through open-source communities like Github and Huggingface, and Tencent Cloud's official website has also launched a model API for convenient and rapid deployment.
If the Hunyuan-A13B model is developed into intelligent document processing applications based on specific business needs within a short time, it significantly lowers the threshold for developers to use the model for secondary development and application innovation.
During the research and development process of Hunyuan-A13B, the Tencent Hunyuan team adopted new technical methods in the pre-training and post-training phases.
In the pre-training phase, a corpus of 200 trillion high-quality web tokens covering multiple fields such as science, technology, and culture was used to enhance the model's general knowledge reserve.
At the same time, the team constructed a Scaling Law joint formula suitable for the MoE architecture, improving the relevant theoretical system and providing quantitative guidance for model architecture design. This achievement serves as an important reference for the subsequent development of MoE models.
In the post-training phase, a multi-stage training approach was adopted, applying different training strategies and data according to various capability enhancement needs; during the inference capability training phase, a large number of logical reasoning case data were used to enhance the model's logical analysis ability; in the creative capability training phase, data from literary creation and copywriting were used to improve the model's text creation level, ultimately balancing the enhancement of the model's reasoning, creation, and understanding abilities.
Tencent Hunyuan simultaneously open-sourced the ArtifactsBench and C3-Bench datasets, filling some gaps in industry evaluation standards.
ArtifactsBench includes 1,825 tasks covering nine major fields such as web development, data visualization, and game development, graded by difficulty, used to evaluate the model's code generation capabilities.
Through this dataset, developers can gain a more comprehensive and accurate understanding of the model's strengths and weaknesses in code writing.
C3-Bench is designed for Agent scenario models, with 1,024 test data points focusing on challenges such as planning tool relationships, handling hidden information, and dynamic path decision-making, helping to identify the model's capability shortcomings in this scenario and providing references for model optimization.
The release of these two datasets provides the industry with more professional and targeted evaluation tools, contributing to the improvement of the large model evaluation system.
Currently, Hunyuan-A13B has been applied in over 400 businesses within Tencent, with an average daily request volume of 130 million times, achieving a certain scale of use in actual business.
For example, in Tencent's intelligent customer service system, the model has improved the accuracy and efficiency of customer service responses; in content creation assistance tools, it helps creators generate higher-quality copy
In the future, Tencent's Hunyuan plan will launch dense models ranging from 0.5B (500 million) to 32B (3.2 billion), as well as activate a MoE model of 13B (1.3 billion) to meet the different needs of enterprises and terminal devices.
At the same time, it will continue to open source multimodal foundational models and plugin models for images, videos, 3D, etc., enriching the large model ecosystem and injecting more vitality into industry development.
Tencent Hunyuan's open sourcing of the Hunyuan-A13B model and related datasets provides developers with new model resources and evaluation tools, which will help promote innovation and application of large model technology.
The release of the open-source dataset also supports the establishment of more comprehensive evaluation standards in the industry. The technical methods developed by Tencent during the R&amp;D process provide reference experience for other teams conducting related research, which is expected to promote the joint development of technology in the large model field

TCEHY

TCTZF

00700

The first hybrid reasoning MoE is open source

- Tencent's Hunyuan team has released the Hunyuan-A13B model, the first open-source Mixture of Experts (MoE) model with 80 billion parameters, enhancing inference efficiency and speed compared to similar architectures.  
- The model supports multi-tool collaboration, enabling complex tasks across various domains, including travel planning and data analysis, while addressing long document processing challenges with a 256K context window.  
- Accompanied by the ArtifactsBench and C3-Bench datasets, this release aims to improve evaluation standards in the industry and has already been integrated into over 400 Tencent applications, demonstrating significant usage and impact.

Track Hyper | Tencent Hunyuan Open Source Hunyuan-A13B: One AI card does it all