AICodeKing evaluates Tencent Hy3 through a hosted endpoint with reasoning set to high. The channel's seven-task KingBench test includes interactive simulations, visual code generation, a counting problem and a multi-step local fine-tuning task.
The reviewer reports incomplete interaction logic, broken controls and failures on the reasoning and agent tasks, for a total of 16 out of 70. These are results from one small, channel-defined benchmark, not a general ranking of models or an independent replication of Tencent's evaluations.
AICodeKing contrasts these observations with the developer's reliability and capability claims. The useful distinction is between a model's advertised characteristics and whether its output actually works for a particular task; different workloads need their own checks.
Watch on YouTube
