Tencent Hunyuan team has announced the release of four open-source small-scale models with parameters of 0.5B, 1.8B, 4B, and 7B. These models are designed for consumer-grade GPUs and are suitable for low-power scenarios such as laptops, smartphones, smart cabins, and smart homes. They support cost-effective fine-tuning in vertical fields. This move further enriches the Hunyuan open-source model system and provides developers and enterprises with more model size options.

The launch of these four models is part of Tencent Hunyuan's continuous open-source efforts, aiming to provide developers and enterprises with more choices to meet different scenario requirements. Currently, these models have been launched on open-source communities such as Github and HuggingFace, and have received support from multiple consumer-grade terminal chip platforms including Arm, Qualcomm, Intel, and MediaTek.

The newly released four models belong to the fused inference models, which are characterized by fast inference speed and high cost-effectiveness. Users can flexibly choose the thinking mode of the model according to their usage scenarios: the fast thinking mode provides concise and efficient output, suitable for simple tasks; the slow thinking mode involves solving complex problems and has more comprehensive reasoning steps. In terms of performance, these models perform well in language understanding, mathematics, and reasoning, achieving leading scores on multiple public test sets.

WeChat screenshot_20250804151331.png

The highlights of these four models lie in their agent capabilities and long-text processing abilities. Through carefully constructed data and a reward signal design for reinforcement learning, these models demonstrate strong performance in agent capabilities such as task planning, tool calling, complex decision-making, and reflection. They can easily handle tasks such as deep searching, Excel operations, and travel plan planning. Additionally, the original long context window of the model reaches 256k, meaning the model can remember and process ultra-long content equivalent to 400,000 Chinese characters or 500,000 English words in one go. It is equivalent to reading three "Harry Potter" novels at once, remembering all character relationships and plot details, and being able to discuss subsequent story developments based on this content.

In terms of deployment, these four models can be deployed with just a single card, and some PCs, smartphones, and tablets can connect directly. The models have strong openness and are supported by mainstream inference frameworks (such as SGLang, vLLM, and TensorRT-LLM) and various quantization formats.

On the application level, these four small-scale models can meet diverse needs from edge devices to the cloud, and from general to specialized scenarios. They have already been applied in multiple business areas of Tencent, and their usability and practicality have been verified through practice. For example, relying on the model's native long-context capability, the AI assistant of Tencent Meeting and the AI Ask Book assistant of WeChat Reading can understand and process complete meeting content and entire books in one go. On edge-side applications, Tencent Mobile Security uses the small-scale model to improve the accuracy of spam SMS identification, achieving millisecond-level interception with zero privacy upload. The Tencent Smart Cabin Assistant solves the pain points of the in-car environment through a dual-model collaboration architecture, fully leveraging the model's low power consumption and efficient inference characteristics.

Official website experience address: https://hunyuan.tencent.com/modelSquare/home/list

[Github]

Hunyuan-0.5B: GitHub - Tencent-Hunyuan/Hunyuan-0.5B

Hunyuan-1.8B: https://github.com/Tencent-Hunyuan/Hunyuan-1.8B

Hunyuan-4B: https://github.com/Tencent-Hunyuan/Hunyuan-4B

Hunyuan-7B: https://github.com/Tencent-Hunyuan/Hunyuan-7B

[HuggingFace]

Hunyuan-0.5B: https://huggingface.co/tencent/Hunyuan-0.5B-Instruct

Hunyuan-1.8B: https://huggingface.co/tencent/Hunyuan-1.8B-Instruct

Hunyuan-4B: https://huggingface.co/tencent/Hunyuan-4B-Instruct

Hunyuan-7B: https://huggingface.co/tencent/Hunyuan-7B-Instruct