Tencent has officially released Hy3, an open-source AI model built around a Mixture-of-Experts architecture. The release adds another large model to the open AI ecosystem, with access through Hugging Face, ModelScope, and GitHub.
The central claim is efficiency at scale. Tencent says Hy3 can match the performance of models two to five times its size, while using only a portion of its parameters at any given time.
A large model that only activates part of itself
Hy3 uses a Mixture-of-Experts, or MoE, design. In plain terms, that means the full model contains many parameters, but not all of them are active for every task.
The model has 295 billion total parameters. Of those, 21 billion are active at any given time. Tencent also adds 3.8 billion parameters for an MTP layer.
That architecture is central to how Tencent frames the model. Rather than presenting Hy3 only as a very large system, the company emphasizes how much of the model is used during operation. The result, according to Tencent, is a model that can compete with larger systems while activating a smaller slice of its total capacity.
Hy3 also supports context lengths up to 256,000 tokens. That gives the model room to process long inputs, which matters for tasks that involve extended documents, large conversations, or complex prompts that require more surrounding information.
The performance claims Tencent is making
Tencent says Hy3 matches the performance of models two to five times its size. That claim is tied to the model's active size rather than only its total parameter count, which is why the MoE design is important to the release.
The company also points to evaluation results. In a blind evaluation by 270 experts, Hy3 scored 2.67 out of 4. GLM-5.1 scored 2.51 in the same comparison.
The source also reports internal testing on hallucinations. In that testing, the hallucination rate dropped from 12.5 percent to 5.4 percent.
Those numbers matter because hallucination remains one of the most important practical issues for AI models. A lower rate, if it holds up in real use, would make the model more useful in settings where users need answers that stay closer to the source material or task context.
Where Hy3 is available
Hy3 is being released under an Apache 2.0 license. That license choice is significant for developers and organizations because it gives the model a clearer path into open-source workflows than a closed release would.
The model is available on several major platforms:
- Hugging Face
- ModelScope
- GitHub
An FP8-quantized version is also available. That gives users another version of the model to consider when working with deployment and resource constraints.
Support for platforms like OpenRouter and Cline is planned. That means access may expand beyond the initial release locations, though the source only describes that support as planned.
How Tencent is already using Hy3
Tencent has already built Hy3 into its own products. The model is being used in WorkBuddy, Yuanbao, WeChat, and the game assistant for "Path of Exile: Advent."
That internal adoption gives the release a practical dimension. Hy3 is not only being published as a standalone model for outside developers; Tencent is also placing it inside products and services it already operates.
For users watching open-source AI, the key point is that Hy3 combines several traits in one release: a large MoE architecture, long-context support, an Apache 2.0 license, public availability, and Tencent's own deployment across multiple products.
What the release signals
Hy3 arrives as another example of how AI developers are trying to balance scale and efficiency. The headline number is 295 billion total parameters, but the active parameter count is 21 billion at any given time. That distinction is the core technical story.
The model's availability also matters. By placing Hy3 on Hugging Face, ModelScope, and GitHub, Tencent is making it easier for developers to inspect, test, and build with the model than if access were limited to a private product.
The claims around expert evaluation and hallucination reduction will likely draw attention from developers comparing open models. Hy3 scored 2.67 out of 4 in a blind evaluation by 270 experts, ahead of GLM-5.1 at 2.51. Internal testing also showed a drop in hallucination rate from 12.5 percent to 5.4 percent.
For now, the release positions Hy3 as an open-source AI model aimed at strong performance without activating its full parameter count for every task. Its broader impact will depend on how developers use it, how planned platform support develops, and how its reported strengths hold up outside Tencent's own tests.