Hy3 is a large language model developed by Tencent, reaching general availability on July 6, 2026 after a preview phase in which it became widely used as a coding assistant on developer platforms. The model's design centers on a Mixture-of-Experts architecture that combines a 295B-parameter total capacity with a much smaller active subset, which independent analyses describe as keeping the model light enough to run on a single multi-GPU node while still projecting a strong overall presence. This sparse-activation approach is the defining technical choice behind Hy3, trading the depth of a fully dense model for the efficiency of selecting only a fraction of the experts per token.
In third-party comparisons with competing open-weight models such as GLM-5.2, Hy3 is positioned primarily as a coding-oriented system aimed at engineers evaluating open-weight large language models for internal deployment. Its preview reception highlighted practical developer value, with reviewers describing it as a go-to assistant for building software more efficiently, and that coding focus carried forward into the general availability release. The combination of a large expert pool, modest active parameters, and a coding-first emphasis makes Hy3 a fit for teams that want an open-weight model capable of substantial reasoning capacity without the compute footprint of a fully dense 295B-parameter model.
