Tencent has officially released the full version of Hy3, its 295-billion-parameter Mixture-of-Experts (MoE) model under the highly permissive Apache 2.0 license. This strategic move eliminates the regional licensing restrictions common to Chinese AI models, allowing global enterprises to adopt it. The release positions Hy3 as a highly competitive open-source alternative capable of challenging rivals like Zhipu AI's GLM-5.2.
Detailed Developments
Following a ten-week preview phase since April, Tencent's Hunyuan team gathered feedback from over 50 internal product teams to refine the model for the official launch. Featuring an MoE architecture with 295B total parameters and 21B active parameters per forward pass, Hy3 is highly optimized for deployment economics. To encourage early adoption and gather community feedback, Tencent is offering the model free on OpenRouter for the first two weeks post-launch.
Technical & Technology Analysis
Architecturally, Hy3 retains its preview specifications with top-8 routing across 192 experts, a 3.8B-parameter multi-token prediction (MTP) layer for speculative decoding, and a 256K context window. The most notable upgrades lie in reliability: the hallucination rate dropped from 12.5% to 5.4%, and commonsense errors fell from 25.4% to 12.7% due to fine-grained data cleaning and training constraints. Notably, Tencent recommends running the model on Nvidia's H20-3e GPUs, which are designed to comply with US export restrictions targeting China.
Expert Opinions & Insights
According to Tencent's blind human evaluation with 270 experts, Hy3 outperformed its older rival GLM-5.1 in real-world workflows like frontend development, CI/CD, and storage. However, experts point out that Zhipu AI's newer GLM-5.2 still retains the coding crown. This performance gap is expected, given that GLM-5.2 is a much larger 744-billion-parameter model with 40 billion active parameters per token, nearly doubling Hy3's compute footprint.
Impact & Future
Tencent's choice of the Apache 2.0 license with no regional restrictions allows Western enterprises to integrate the model without legal hurdles. By offering frontier-adjacent performance at half the size of GLM-5.2, Hy3 drastically reduces infrastructure costs, making self-hosting far more accessible. It serves as an highly cost-effective option for organizations looking to build search-and-tool-heavy agent workflows without paying for massive, enterprise-scale hardware cluster requirements.