Tencent released the source code of the new language model Hunyuan-A13B, which can dynamically switch between fast and deep “thinking” depending on the complexity of the task. The model allows users to change the depth of reasoning in real-time: for simple queries, a fast mode is used, and for complex ones, multi-step processing is activated. This can be controlled using special commands — “/think” for deep thinking and “/no_think” for normal mode.
🚀 Introducing Hunyuan-A13B, our latest open-source LLM.
— Hunyuan (@TencentHunyuan) June 27, 2025
As an MoE model, it leverages 80B total parameters with just 13B active, delivering powerful performance that scores on par with o1 and DeepSeek across multiple mainstream benchmarks.
Hunyuan-A13B features a hybrid… pic.twitter.com/8QTT547fcC
Hunyuan-A13B is built on the Mixture of Experts architecture with a total of 80 billion parameters, but only 13 billion are active during operation. The model supports large context windows — up to 256,000 tokens, allowing it to work with extensive texts and tasks. For training, 20 trillion tokens were used, of which 250 billion were collected from STEM fields, including math textbooks, tests, open code from GitHub, and scientific texts of various levels.
The model performs well in scientific and mathematical tasks. In the AIME 2024 math competition, it achieved an accuracy of 87.3 percent, surpassing OpenAI o1 in the same round. Internal Tencent tests showed that Hunyuan-A13B maintains high results on many agent tasks and demonstrates stability even when working with large contexts, although it falls short of Gemini 2.5 Pro in some tests.
Hunyuan-A13B is already available under the Apache 2.0 license on Hugging Face and GitHub. For quick deployment, Docker images are prepared, and there is access through the API in Tencent Cloud and a browser demo version. Tencent also added two new datasets for testing: ArtifactsBench for code generation and C3-Bench for agent task evaluation.
