The Chinese technology giant scaled up the infrastructure for its WorkBuddy AI platform this week to address usage limits caused by a spike in traffic following the release of its new Hy4 foundation model. To manage the initial volume, the company increased the capacity of its inference clusters by 10 times and extended free access to its previous model, Hy3, through Nov 15. The Hy4 model is now available for use on Tencent Cloud, OpenRouter, and OpenCode Go, with certain API services offering a 50% discount until the end of 2026.
The new model supports a 1M token context window and was specifically trained for the production of game development software and the editing of office documents. While Hy4 recorded a 31% score on the FrontierCode 2.0 benchmark to lead all other models from China-based labs in agentic coding, it struggled with accuracy, recording an 89% hallucination rate on the AA-Omniscience v1.1 metric. Tencent plans to open-source the weights for the flagship model and release a lower-cost “Flash” version in the coming weeks.
