Tencent open-sources Hy4 preview for long-context productivity

Tencent released and open-sourced Hy4 preview on August 28, 2026, positioning the large language model for coding, office work and scientific research. The company says Hy4 has 770 billion total parameters, of which 49 billion are active during inference, and supports a context window exceeding one million tokens. The model is available through Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima, through Tencent Cloud TokenHub and OpenRouter, and as open-source software.

The launch matters because it combines a very large mixture-of-experts model with routes for both local or independently managed deployment and hosted use. Developers can evaluate the model directly instead of relying only on a closed application. Tencent is also offering two weeks of free Hy4 preview access in WorkBuddy and CodeBuddy, giving users an immediate way to test its behavior on software and document workflows. Exact hardware needs, license conditions and operating costs will remain important practical considerations for teams planning their own deployment.

Tencent reports that the model was trained with data created alongside specialists in software engineering, gaming, finance and security. In an internal blind evaluation involving 163 experts and 203 engineering tasks, Hy4 preview scored 2.99 out of four, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Those figures are Tencent's own results, not an independent benchmark, and the comparison does not by itself establish broader model quality.

The company also says Hy4 participated in parts of its own development process, proposing and testing changes to training methods, data strategies, evaluations and low-level operators. According to Tencent, model-assisted inference optimizations improved end-to-end throughput by 31.8 percent against its baseline. This is best read as a reported engineering result rather than proof of autonomous recursive improvement.

For AI makers and businesses, Hy4 adds another heavyweight open model to a fast-moving market increasingly shaped by Chinese labs. Its long context and productivity focus could be useful for large codebases, multi-document analysis and research workloads. The open release also gives researchers more room to inspect, adapt and compare the system. Real-world value will depend on independent tests of quality, reliability, safety, memory use and total deployment cost.