Hy3 is Tencent's flagship open-source model designed for enterprise agentic workflows, built on a Mixture-of-Experts architecture where only 21 billion of its 295 billion total parameters activate per query. This selective activation keeps inference efficient while preserving access to specialised capabilities across diverse tasks. The model natively handles a 256,000-token context window, comfortably accommodating long documents, extended call transcripts, and multi-turn agentic reasoning chains without truncation. Tencent refined Hy3 over several months using feedback from 50 internal product teams and real-world deployment data from Yuanbao, WorkBuddy, ima, and Marvis, focusing improvements on coding agent performance, long-form understanding, multi-turn context continuity, search QA, and complex task execution.
Hy3 offers three thinking modes that let users trade response speed against reasoning depth: a no-think ultra-fast mode, a think-low quick reasoning mode, and a think-high deep reasoning mode. This flexibility makes the model well-suited to practical enterprise scenarios including frontend development, cross-file code work, long-document analysis, office automation, and multi-step agent pipelines. Tencent reports that hallucinations dropped from 12.5% to 5.4% and commonsense reasoning errors fell from 25.4% to lower levels between the preview and full release, reflecting the impact of that production-driven refinement loop. Released under an Apache 2.0 license, Hy3 is positioned as a commercially friendly option for organisations weighing open-source flexibility against frontier-tier agent capabilities.
