Tencent released and open-sourced Hy4 preview on August 28, 2026, describing it as the next Hunyuan large language model. The company says the preview has 770 billion total parameters, 49 billion active parameters, and a context window of more than one million tokens. Tencent is positioning the release around coding, office work, and scientific research rather than a general chatbot demo.

The weights are available as an open-source download. The model also ships through Tencent’s WorkBuddy and CodeBuddy products, the Yuanbao and ima apps, and APIs on Tencent Cloud TokenHub and OpenRouter. WorkBuddy and CodeBuddy will offer Hy4 preview for free for two weeks. Free access to the earlier Hy3 model on those apps is extended through September 30.

Productivity claims and a self-tuning loop

Tencent says Hy4 preview was trained on data built with its own specialists in software engineering, gaming, finance, and security, then tuned against WorkBuddy and CodeBuddy. In an internal blind test with 163 experts and 203 engineering tasks, the company scored Hy4 preview at 2.99 out of 4.00, just ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94.

The announcement walks through four jobs. In software engineering, Tencent claims stronger planning and debugging on long-context work plus better front-end visuals. In office settings, it highlights financial analysis, data work, and turning source material into documents, spreadsheets, and slides. In games, it says a single natural-language prompt can produce a playable prototype that developers then refine. In research, it cites gains in molecular dynamics, condensed-matter physics, and fundamental mathematics.

Tencent also says Hy4 preview helped design later training methods, data mixes, evaluation harnesses, and low-level operators, then used that output in the next round. The company reports that the model found inference bottlenecks and raised end-to-end throughput 31.8 percent versus its own baseline through operator fusion and communication changes. API list prices are $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens on cache hits. Hunyuan says a fuller Hy4 family will follow the preview.

Decoded Take

A 770 billion parameter open release from Tencent is less about one leaderboard and more about keeping Chinese weights in the same conversation as GLM and Kimi for people who will actually host a model. The interesting claim is not the 2.99 score. It is the assertion that the model already wrote parts of its own training stack and sped up its own serving path. That is still a company-run story, not an outside audit. The signal to watch is whether third-party downloads on TokenHub and OpenRouter show Hy4 preview being used for long-context coding and document work after the two-week free window closes, or whether it stays a preview that never becomes the default Hunyuan checkpoint.