Tencent Hunyuan-A13B is online and can be deployed with one mid- to low-end GPU card

On June 27, Tencent publicly released a new member of the Hunyuan model family - the Hunyuan-A13B model, and announced that it will be fully open source. According to official sources, the Hunyuan-A13B model can be deployed under extreme conditions with only one mid- to low-end GPU card.

In the mathematical reasoning experiment, the model accurately completed decimal comparisons and demonstrated step-by-step analysis capabilities. It took 3 seconds to think deeply.

The Hunyuan-A13B model also made technological innovations. In pre-training, the model used a 20 trillion high-quality network word corpus, and improved the Scaling Law theoretical system of the MoE architecture, providing quantifiable engineering guidance for the design of the MoE architecture.

Hunyuan has also open sourced two new data sets to fill the gaps in relevant evaluation standards in the industry. Among them, ArtifactsBench is mainly used for code evaluation and builds a new benchmark containing 1825 tasks; C3-Bench designs 1024 test data for Agent scene model evaluation to discover deficiencies in model capabilities.

Currently, the model API is online on the Tencent Cloud official website.