Tencent's Hy team has unveiled 'Hy3,' a large-scale open AI model. Hy3 is a language model with a Mixture-of-Experts (MoE) structure featuring a total of 295 billion parameters, with 21 billion parameters actually active per token. While the overall model is massive, it is characterized by improved computational efficiency through a method that selectively uses only a subset of expert models during inference. Licensed under Apache 2.0, it offers relatively open conditions for both research and commercial use.

Hy3 supports a token context window of up to 256K, enabling it to be utilized for long document analysis, understanding large codebases, complex multi-turn conversations, and agent tasks requiring long-term memory. Tencent explained that the model was designed with a focus on inference, agentic workflows, and long-text context processing. As the recent competition in AI models expands beyond simple parameter scale to actual active parameters, context length, tooling capabilities, and hallucination reduction performance, Hy3 represents an example of expanding high-performance MoE model options within the open model ecosystem.

According to released performance metrics, Hy3 scored 78.0 points on SWE-Bench Verified, a software engineering benchmark. Furthermore, by reporting a reduction in hallucination rates, it is drawing attention for its potential applications in fields where reliability is critical, such as enterprise AI, developer assistance tools, and research automation. Hy3 is available for free trial on OpenRouter until July 21, 2026. By offering an open license, processing 256,000 long texts, and efficiency based on MoE, Hy3 is emerging as a key choice for developers and enterprises looking to build large-scale language models or integrate them into their services.