Tencent Hunyuan launches another Open Source AI big move! The Hunyuan-A13B model debuts, challenging the 80B giant with 13B parameters.

As generative AI enters a new phase, while the super-large models are powerful, they also come with high resource costs and deployment difficulties. Tencent's latest release, the open source large language model Hunyuan-A13B, has taken a lightweight and efficient new approach: although it has a total of 80 billion parameters, only 13 billion parameters are activated for each inference. Through the "Mixture-of-Experts (MoE)" architecture, it significantly reduces inference costs without sacrificing language understanding and reasoning capabilities.

Breaking the Performance and Resource Bottleneck: Technical Underpinnings of Hunyuan-A13B

Hunyuan-A13B adopts a sparse MoE architecture, with a total of 64 expert modules and one shared expert. During each inference, only 8 experts are activated, combined with the SwiGLU activation function and GQA (Grouped Query Attention) technology, effectively improving memory usage efficiency and inference throughput.

In addition, its pre-training data amounts to as much as 20 trillion tokens, including 25 billion high-quality data from the STEM field, enhancing the model's performance in mathematics, science, and logical reasoning. The overall training goes through three major stages: basic pre-training, rapid annealing training, and long context expansion training, ultimately supporting a context processing capability of up to 256K tokens.

Choose between two inference modes: Quick vs. Deep Thinking freely switch.

Hunyuan-A13B specially introduces a dual-mode inference mechanism (Dual-mode Chain-of-Thought), which automatically switches between "fast thinking" and "slow thinking" modes based on task complexity. Fast thinking is suitable for simple daily queries, emphasizing low latency and high efficiency; slow thinking activates a longer Chain-of-Thought reasoning path, handling multi-step logic and complex inferential problems, balancing accuracy and resource usage.

The evaluation results are impressive.

In multiple recognized benchmark tests, the Hunyuan-A13B performed exceptionally well. It topped the charts in mathematical reasoning (such as AIME 2024, MATH) and also ranked high in logical reasoning (such as BBH, ZebraLogic). Even in long-text tests that challenge model depth, such as LongBench-v2 and RULER, the Hunyuan-A13B demonstrated impressive context retention and logical integration capabilities, defeating Qwen3-A22B and DeepSeek-R1, which have far more parameters than it.

In the face of challenges in an open environment, the tool invocation capability has been fully upgraded.

In addition to language and reasoning capabilities, Hunyuan-A13B has also significantly improved its Agent capabilities. In tests such as BFCL v3 and ComplexFuncBench, which focus on tool invocation, planning, and multi-turn dialogue, its performance not only surpasses Qwen3-A22B but also ranks first in the C3-Bench test, demonstrating its strong ability to adapt to complex task processes.

The inference throughput is off the charts, and the deployment efficiency is worth paying attention to.

According to the report, the Hunyuan-A13B is also impressive in reasoning efficiency. When paired with existing reasoning frameworks such as vLLM and TensorRT-LLM, it can achieve a throughput of nearly 2000 tokens/s with a batch size of 32 and an output length of 14K tokens. Even under quantization precision conditions like INT8 and FP8, it can maintain its performance, which is crucial for enterprise deployment.

High performance and high cost-effectiveness are both achieved, Open Source models welcome new choices.

Tencent sets a new benchmark for the open source language model community with Hunyuan-A13B. This model not only breaks the stereotype of "small models being no match for large models", but also offers flexible reasoning and multi-task adaptability, becoming a new flagship in the open source LLM field. For developers and enterprises with limited resources but still pursuing efficient AI solutions, Hunyuan-A13B is undoubtedly a powerful new option worth attention.

This article Tencent Hunyuan launches another Open Source AI big move! The Hunyuan-A13B model is unveiled, challenging the 80B behemoth with 13B parameters, first appearing in Chain News ABMedia.