Hunyuan-A13B

What it is

Hunyuan-A13B is a large language model from Tencent's Hunyuan team, released as open weights, which means its trained parameters can be downloaded and run by others. It is a fine-grained mixture-of-experts (MoE) model with 80 billion total parameters, and 13 billion of them are active for each token it processes.^[1] Activating only part of the network on each step lets a model be large in total while doing less work per token than a dense model of the same total size.

Origins

Tencent released the model on Hugging Face on 27 June 2025, together with a technical report and a training and inference manual.^[1] Several variants were published at once: a pretrained base model (Hunyuan-A13B-Pretrain), an instruction-tuned model (Hunyuan-A13B-Instruct), an FP8 static-quantized Instruct model and a GPTQ-Int4 (W4A16) Instruct model.^[1] Quantization stores a model's weights at lower numerical precision, which reduces the memory needed to run it.

Main features

The model supports a native 256K-token context window.^[1] It uses grouped-query attention, a variant of the attention mechanism in which groups of query heads share key and value heads, which reduces memory use during inference.^[1]

Its reasoning is hybrid. Slow, chain-of-thought thinking is the default, and users can turn it off with the setting enable_thinking=False or by adding a /no_think prefix to a prompt.^[1]

Tencent reported the following results for the Instruct model: 87.3 on AIME 2024, 76.8 on AIME 2025, 94.3 on MATH and 71.2 on GPQA-Diamond.^[1] It also highlighted agent-oriented tests, with 78.3 on BFCL v3, 54.7 on τ-Bench and 63.5 on C3-Bench.^[1] For the base model, the reported figures were 88.17 on MMLU, 67.23 on MMLU-Pro and 91.83 on GSM8K.^[1] The FP8 and Int4 quantized versions both scored 86.7 on AIME 2024.^[1]

The model can be deployed with TensorRT-LLM, vLLM and SGLang.^[1]

Place in the field

Hunyuan-A13B shows Tencent publishing a mid-sized mixture-of-experts model with its weights, a technical report and a deployment guide together. Because of this, outside developers could examine its behaviour, run it on their own hardware and choose among the released precisions.