Tencent's Hy3 Model Released Under Permissive License, Rivals Top Open-Weight AI
Tencent has released the full version of its Hy3 large language model under the permissive Apache 2.0 license, removing previous restrictions that limited its use in regions like the EU, UK, and South Korea. This move is significant for enterprises, as it makes a powerful Chinese AI model legally accessible for deployment. Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) model with 21 billion active parameters, was developed using Tencent's rebuilt pre-training and reinforcement learning infrastructure over ten weeks, incorporating feedback from over 50 internal product teams. The model boasts a 256K context window and aims to outperform similar-sized models while being competitive with flagship open-source models that have two to five times its parameters.
Tencent's evaluation highlights Hy3's strengths in reliability and production economics, reporting halved hallucination and commonsense error rates compared to its preview version. In internal tests, hallucination rates dropped from 12.5% to 5.4%, and commonsense error rates fell from 25.4% to 12.7%. The model also shows improved multi-turn conversation capabilities, with issue rates decreasing from 17.4% to 7.9% in internal tests. While Hy3 demonstrates strong performance in agentic search and tool orchestration, Tencent's own benchmarks indicate that Zhipu AI's GLM-5.2 model, which has significantly more parameters, retains the lead in coding tasks.
The deployment economics of Hy3 are also a key selling point, with its 295B parameters requiring less than half the memory of GLM-5.2's 744B parameters. This makes Hy3 more attainable for self-hosting, fitting within more accessible hardware configurations. Notably, Tencent's recommended serving configuration targets Nvidia's H20-3e GPUs, designed to comply with U.S. export restrictions, suggesting the model is optimized for hardware legally available to Chinese companies. This design choice also benefits Western data centers, as a model optimized for constrained silicon performs even better on higher-end GPUs.
Tencent's strategic release of Hy3 under the Apache 2.0 license addresses a critical market friction point: the accessibility of advanced AI models for global enterprise deployment. By shifting from restrictive licenses to a permissive one, Tencent not only broadens its potential user base but also signals a commitment to open-source principles, potentially reshaping competitive dynamics in the open-weight model landscape. The model's architecture, emphasizing efficiency with fewer active parameters than some rivals, suggests a design philosophy prioritizing deployment economics and performance on hardware with export controls. This focus on reliability metrics and reduced hallucination rates, alongside optimized resource utilization, indicates a pragmatic approach to production readiness, catering to enterprises wary of the operational costs and risks associated with less stable models. The trade-off in coding performance against larger models like GLM-5.2 highlights a deliberate strategic choice, prioritizing broader applicability and cost-effectiveness over niche, albeit important, capabilities. This positions Hy3 as a compelling option for organizations seeking a balance between performance, accessibility, and operational efficiency in the rapidly evolving AI ecosystem.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.