Tencent's hunyuan 4 leverages open source and talent to challenge large model leaders

Tencent's Hunyuan 4 (Hy4 preview) has demonstrated how open-source technologies and top talent can dismantle perceived technical barriers in large models. Yao Shunyu, who previously worked at OpenAI, stated that there are no real secrets to building large models, a claim that gained credibility after Tencent's recent release of Hy4. The model, with 770B parameters and support for a million context tokens, leverages open-source innovations from companies like Zhipu and DeepSeek. It integrates sparse attention mechanisms and cross-layer index reuse, significantly improving performance. Bai Yushi, a former Zhipu researcher, joined Tencent's team, facilitating the transfer of expertise. The rapid dissemination of cutting-edge technologies through open-source code and community discussions is eroding industry advantages, as noted by industry observers. Hy4's architecture aligns with Zhipu's GLM-5, adopting a hidden dimension of 6144 and 78 layers. It also incorporates Microsoft researcher Xie Tian's identity hyperconnection (iHC) for residual optimization. These technical choices, combined with talent mobility, highlight a shift in the large model landscape. As open-source ecosystems grow, the competitive edge of large models is being rapidly diminished, according to industry experts.