Tencent's Hy4 AI model beats rivals in coding tests

Tencent quietly released one of the largest open-source AI models to come out of China this year, and the company didn't wait to compare it against the competition.

Hy4 preview, published Friday on Hugging Face, scored 2.99 out of 4.00 in Tencent's own blind evaluation across 163 experts and 203 engineering tasks, narrowly ahead of Zhipu's GLM-5.3 at 2.92 and Moonshot's Kimi K3 at 2.94.

Hy4 preview carries 770 billion total parameters, though only about 49 billion are activated for any single request, using a mixture-of-experts design that routes each query through a smaller subset of the full network.

That's more than double the scale of Hy3, Tencent's previous flagship, which launched in April with 295 billion total parameters and 21 billion active ones.

The new model also extends its context window past 1 million tokens, letting it process far more text in a single request than its predecessor could.

Unlike other language models that use benchmark datasets for training, Hy4 has been trained by Tencent using datasets generated by the company's software engineers, video game developers, financial analysts, and security experts in their daily work.

According to the company, the preview version of Hy4 can take a lot of time to process complicated enquiries and can sometimes unnecessarily double-check the correct answer.

However, Tencent is going to fix both problems in the future versions of the language model.