Tencent has released its most capable AI model yet, Hy3 preview, which went open-source across GitHub, Hugging Face, and ModelScope. The model packs 295 billion total parameters but only 21 billion active at any given time. It uses a Mixture-of-Experts architecture, which routes queries to a specialized subset of its "expert" sub-networks, and supports 256,000 tokens of context. The previous model, Hy2, had over 400 billion parameters, but this model was designed to balance capability breadth, honest evaluation, and cost-efficiency. Hy3 still trails in the industry by comparison with other AI models. On a benchmark tracking complex web research tasks, it reached 67.1% (up from Hy2's 28.7%) and 77.2%. In reasoning, the model topped all Chinese competitors on Tsinghua University's math PhD qualifying exam (Spring 2026), scoring 88.4 on the average of three runs avg@3.