Tencent Hy4: 770B-Parameter Open-Weight AI Model Adds Vision

Bigger model, more to come

Hy4 has 770 billion parameters, about 2.6 times the 295 billion of its predecessor, Hy3. The developer says the next generation, Hy5, will be larger still, and a lighter Hy4 Flash variant is due soon to cut the cost of running tasks.

Connected to WorkBuddy, the model reached an average score of 83.18 with a 96.15% success rate, according to Tencent's internal evaluations.

A preview version of Hy4, shown in late August, currently ranks 23rd on the WebDev leaderboard of Code Arena, which scores language models in real time on practical programming tasks. Thanks to far more training data and a larger model, Hy4 beats its predecessor in 12 categories, including autonomous action, coding and workflow automation. It outperformed Z.ai GLM-5.3 and DeepSeek-V4.1-Flash-1017 and performed on par with Moonshot Kimi K3.1 Pro.

Limits and safety choices

Tencent acknowledges that Hy4 still trails leading closed systems, particularly on hard tasks such as mathematical research, long-running coding projects, self-improvement and advanced visual analysis, including answering questions about images.

During training, the company also curbed hacking-related capabilities. Material on exploit development, reverse engineering and penetration testing was removed from the training data to avoid boosting offensive cyberattack abilities.

For developers and businesses, open weights mean a large multimodal model they can inspect and deploy themselves, while the planned Flash version should address running costs.