According to Perceiving AI monitoring, Tencent Quantum officially announced the Hy3 preview, confirming that the model weights and code have been open-sourced on four platforms: GitHub, HuggingFace, ModelScope, and GitCode, supporting mainstream inference frameworks such as vLLM and SGLang. Tencent Cloud simultaneously launched API services and customized Token Plan packages, with the Personal Plan starting at a minimum of 28 RMB per month.
The officials stated that the code and agent are the most significant aspects of the Hy3 preview, providing benchmark data such as SWE-Bench Verified, Terminal-Bench 2.0, ClawEval, WildClawBench, among others. On the CodeBuddy and WorkBuddy products, the first token latency has been reduced by 54%, end-to-end duration decreased by 47%, the success rate has exceeded 99.99%, and it has stably operated the most complex agent workflow for up to 495 steps. Tencent mentioned that this is a preview version and hopes to collect community feedback through open-sourcing to enhance the official release.
