Tencent Unveils Hy4, a 770B‑Parameter Open Model Built to Keep Improving Itself

Chinese tech giant Tencent has previewed a massive open‑weight language model called Hy4, with a reported 770 billion parameters—putting it in the same size class as the largest commercial systems, according to recent AI news roundups. Tencent pitches Hy4 as a “self‑improving” model, designed to ingest user feedback and new data to continuously refine its behavior over time.

A key selling point is its enormous context window, large enough to process the equivalent of an entire novel in one go. That opens the door to use cases like reasoning across whole codebases, analyzing months of project documentation, or performing complex legal and financial reviews that span hundreds of pages—tasks where conventional models often lose track of earlier details. Early reports also point to experimental agent‑like capabilities, allowing Hy4 to optimize prompts and choose tools autonomously during inference.

Tencent has long offered proprietary AI services on its cloud, but releasing a frontier‑scale model as open weights marks a more aggressive play against open‑source ecosystems built around Meta’s Llama or Chinese rival DeepSeek. The move also highlights geopolitical constraints: training and serving a 770B‑parameter model requires immense GPU capacity at a time when US export controls are tightening. Whether developers and researchers embrace Hy4 will be a key indicator of Tencent’s ability to build a global AI ecosystem under those hardware and regulatory pressures.