Tencent Holdings Ltd. today open-sourced a new lineup of language models, the Hunyuan-MT series, that is optimized for translation tasks.

The Chinese tech firm says that the algorithms significantly outperform Google Translate on a popular artificial intelligence translation benchmark.

The Hunyuan-MT series comprises four models. The two flagship models, Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B, both feature 7 billion parameters. There are also two quantized, or compressed, versions of the algorithms that trade off some output quality for lower memory usage.

Tencent carried out the models’ initial training using four different datasets. The first two contain text snippets written in 33 languages, but don’t include any information on how to perform translation. The other two datasets include several million so-called pairs. Those are records that each comprise a text snippet and a translation.

According to Tencent, the training workflow equipped its Hunyuan-MT models with not only translation capabilities but also a significant amount of general knowledge. The company put the algorithms to the test using a general knowledge benchmark called MMLU-Pro. Hunyuan-MT outperformed Llama-3-8B-Base, a model with 8 billion parameters, by a significant margin.

Tencent followed up the initial training with a reinforcement learning stage. In that part of the project, the company gave Hunyuan-MT models a series of training tasks and provided feedback on the quality of their responses. The models used this feedback to improve their output quality.

The trial and error learning process was supervised by a custom AI model. The model scored translations generated by Hunyuan-MT based on their semantic similarity to the original text. It also considered certain other factors, including how well the algorithms processed domain-specific terminology.

The first AI in the Hunyuan-MT series, Hunyuan-MT-7B, is based on a standard language model architecture. Hunyuan-MT-Chimera-7B uses a more complicated processing approach known as ensemble learning. 

Similarly to mixture-of-expert models, an ensemble learning algorithm comprises multiple neural networks. But whereas a mixture-of-expert model uses only one of its neural networks to process the user’s prompt, ensemble learning uses all of them. It generates multiple answers to the prompt and then combines them into a single higher-quality response.

Tencent compared Hunyuan-MT with Google Translate using an AI translation benchmark called WMT25. According to the company, its model series performed better across 30 out of 31 language pairs evaluated in the test. In some cases, Hunyuan-MT scored 65% higher.

The model series also outperformed algorithms from several other AI providers. Tencent says that Hunyuan-MT scored higher than GPT-4.1 and Anthropic PBC’s Claude 4 Sonnet across most of the language pairs in the WMT25 benchmark.

Image: Unsplash