Tencent releases 770B Hy4 preview model

Hy4’s real bet is practical, persistent work rather than benchmark theater—but its 1.56 TB footprint means “open” still largely favors hosted APIs and serious infrastructure teams. [GitHub](https://github.com/Tencent-Hunyuan/Hy4-preview)

• –1M context and native reasoning/tool-call parsers make it well-suited to long-running coding agents and document workflows. [GitHub](https://github.com/Tencent-Hunyuan/Hy4-preview)
• –Apache 2.0 licensing gives developers broad freedom to fine-tune, deploy, and integrate the model commercially. [GitHub](https://github.com/Tencent-Hunyuan/Hy4-preview)
• –Tencent’s internal blind evaluation scored Hy4 at 2.99/4.00, narrowly ahead of GLM-5.3 and Kimi K3, but independent validation is still needed. [Tencent](https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/)
• –The 1.56 TB model size and recommended eight-way tensor parallel deployment make local experimentation expensive despite the open weights. [Hugging Face](https://huggingface.co/tencent/Hy4-preview)