The model is relatively small with only 295 billion parameters, bucking a recent trend of large models with trillions of parameters
The Shenzhen-based company said on Thursday that the closed-source HY3-Preview was its most powerful model yet, on par with leading Chinese models, while still lagging top US models from the likes of OpenAI and Google DeepMind.
Notably, the model is relatively small with only 295 billion parameters, bucking a recent trend of large models with trillions of parameters – the mathematical variables that encode a model’s “intelligence”, which are roughly proportional to the amount of computational power needed to train and serve the model.
The company’s previous flagship model, the HY 2.0 released in early December, had over 400 billion parameters.
HY3-Preview was developed to cater for real-world business scenarios, the company said, highlighting the collaboration between Tencent’s foundational model development team Hunyuan and its Yuanbao AI application team.
“By seamlessly aligning product-side requirements with underlying technology, we have successfully bridged the gap between model capability and user value,” it said, adding that the model was already deployed in Tencent’s flagship AI products including consumer app Yuanbao and coding assistant CodeBuddy.
Source: News - South China Morning Post
