Tencent Open-Sources Hy4, a 770B-Parameter AI Model Built for Science and Productivity

Tencent has open-sourced Hy4 preview, a new Mixture-of-Experts large language model designed for coding, office work, data analysis, and scientific research. The model contains 770 billion total parameters, with 49 billion activated per token, and supports a context window of approximately 1 million tokens. That gives it the ability to work across very large documents, codebases, and research materials in a single workflow, according to Tencent’s announcement.

The release matters because Tencent is positioning Hy4 as a practical productivity system rather than simply a benchmark project. Tencent says the model was developed with input from software engineers, game developers, finance analysts, and security experts. Its intended uses include debugging complex projects, creating documents and presentations, analyzing financial information, and assisting with research in areas such as mathematics, physics, molecular dynamics, and artificial intelligence.

For everyday users, the most notable feature may be the enormous context window. A model that can process information spread across many files could help researchers, developers, and business teams reduce the time spent sorting and summarizing source material. Hy4 preview and its FP8 version are also available through Hugging Face, with deployment support for tools including vLLM and SGLang.

Tencent also reports that Hy4 helped optimize parts of its own inference system, including operator fusion and communication processes. Those changes reportedly increased end-to-end throughput by 31.8% compared with a baseline. The company describes this as an early form of self-optimizing inference: not a fully independent AI, but a model contributing to improvements in the systems that run it.

As with any preview model, users should verify important outputs and avoid placing sensitive information into untrusted systems. Follow Brownstone Worldwide’s technology coverage for updates as Hy4 moves from preview testing into broader use. Would a million-token context window change how you approach research or software development?