Tencent has released the full version of its Hy3 model, a 295-billion-parameter Mixture-of-Experts (MoE) model, under the permissive Apache 2.0 license, removing previous geographical restrictions that hindered its adoption. This license change, a significant reversal from its preview release, has been widely applauded by the open-model community, with some suggesting it positions Tencent as a leader in open source.Hy3's architecture remains consistent with its April preview, featuring 295 billion total parameters and 21 billion active parameters per forward pass. Tencent claims the full release substantially outperforms similar-sized models and competes with flagship open-source models that are two to five times larger. The model was shaped by feedback from over 50 internal product teams, leading to improvements in task execution and interaction.While Tencent's blind human study indicates Hy3's superiority over GLM-5.1 in certain development areas, it trails GLM-5.2 in agentic coding benchmarks. GLM-5.2, being a larger model with roughly 744 billion parameters, maintains its lead in coding performance. However, Hy3 excels in agentic search, tool orchestration, and long-context retrieval, making it a strong contender for search and tool-heavy agent workloads.A key focus of Hy3's release is its emphasis on reliability metrics and deployment economics, rather than solely benchmark scores. Tencent reports a significant reduction in hallucination rates and commonsense error rates compared to its preview version, attributing these improvements to fine-grained data cleaning and training constraints. These reliability improvements address a critical concern for enterprise buyers seeking models that perform consistently in production environments.From an economic standpoint, Hy3 offers a more attainable deployment solution than GLM-5.2. With an FP8 footprint of under 300GB, Hy3 requires less memory and compute, reducing the hardware barrier for self-hosting. Tencent’s recommended serving configuration targets Nvidia’s H20-3e, a GPU designed for U.S. export restrictions on China, ensuring broader accessibility. This constraint-driven design makes the model efficient on various hardware, including those available in Western data centers.The combination of the permissive Apache 2.0 license, strong performance in specific agentic tasks, enhanced reliability, and more accessible deployment economics positions Hy3 as a viable option for enterprises. It particularly appeals to organizations prioritizing reliability and cost-effective infrastructure over absolute coding performance. The remaining uncertainty lies in whether Western enterprises will embrace a Tencent model given its Chinese origin, and how future independent benchmarks will impact adoption.