Tencent open-sources Hy4, a 770-billion-parameter AI model for coding and research

Tencent has released Hy4, a huge new open-source AI model aimed at coding, research and office work. The company says the preview packs 770 billion total parameters into a mixture-of-experts design, although only 49 billion are activated for each token. That makes it one of the more ambitious open models to land this year — provided your idea of “running locally” involves a small data centre.

What’s actually going on: Tencent’s Hy Team describes Hy4 as a model with a one-million-token context window, 78 layers, 256 routed experts and eight experts activated per token. The preview and an FP8 version are available under the Apache 2.0 licence, with official links to Hugging Face, ModelScope, GitCode and CNB. Tencent positions it for software engineering, office and financial analysis, game development and scientific research.

The official repository also supplies a useful dose of self-awareness. Tencent says this early version can spend too long reasoning and may over-verify its answers. The team presents the model’s benchmark results as internal evaluations involving 163 Tencent experts, so those figures should be treated as company-reported results rather than independent proof that Hy4 outperforms every rival.

Why you should care: open weights matter because developers can inspect, adapt and deploy a model instead of relying entirely on a hosted chatbot. A one-million-token context window could also make large codebases and lengthy research material easier to handle. The catch is practical rather than glamorous: Tencent has not published independently verified hardware requirements in the material checked, and a model this large is unlikely to be a casual laptop download.

Our opinion

Hy4 is an intriguing release, especially under Apache 2.0, but the headline parameter count is not the same thing as proven performance. The sensible move is to watch independent testing before rebuilding your workflow around it. For researchers and developers with serious hardware, this is a model worth experimenting with; for everyone else, the open licence is more immediately interesting than the prospect of running 770 billion parameters beside the kettle.