Tencent has officially released the full weights for Hy3, a massive 295-billion-parameter Mixture-of-Experts (MoE) model, under the permissive Apache License 2.0. This launch follows the initial Hy3-preview released earlier this year on April 23. The model features 295 billion total parameters, utilizing 192 experts with top-8 routing, activating only 21 billion parameters per token during inference. The package includes standard BF16 weights alongside FP8 variants and incorporates 3.8 billion MTP layer parameters. Tencent designed Hy3 for complex STEM reasoning and backend coding tasks. According to the official model card, the April preview achieved a 79.26 MMMLU 5-shot score, and the new July release features a massive 256K context window. However, all benchmark figures are self-reported vendor claims, and independent evaluators such as Epoch AI have not yet verified the data.