Hy4 preview powers playable games in WorkBuddy

The impressive part is not the parameter count—it is testing frontier-scale models on long-horizon software delivery, where agents must plan, code, debug, and iterate toward something playable.

• –Apache 2.0 weights make Hy4 preview broadly usable, though its roughly datacenter-scale deployment requirements limit local experimentation. [Official GitHub release](https://github.com/Tencent-Hunyuan/Hy4-preview)
• –WorkBuddy turns game generation into a practical demonstration of multi-step coding, visual inspection, and refinement rather than a static code-generation benchmark.
• –The 1M-token context window should help with larger, multi-file projects and extended iteration, while the 49B active-parameter figure does not make serving the full model lightweight.
• –Tencent’s internal evaluation slightly beats GLM 5.3 and Kimi K3, but those results are vendor-reported and should be treated as an early signal, not definitive proof of superiority.
• –Two weeks of free access in WorkBuddy lowers the barrier for developers to test whether the model can reliably finish real projects. [Tencent announcement](https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/)