Tencent’s Hy3 Open-Source AI Model Claims Efficiency Gains with MoE Architecture

Tencent has officially launched Hy3, a new open-source AI model designed to deliver high performance with enhanced efficiency. The model incorporates a Mixture-of-Experts (MoE) architecture, featuring a substantial 295 billion total parameters, though only 21 billion of these are actively engaged at any given time during processing. An additional 3.8 billion parameters are dedicated to an MTP layer, contributing to its overall capability.

This strategic design allows Hy3 to handle extensive context lengths, supporting up to 256,000 tokens. Tencent asserts that Hy3’s performance is on par with models two to five times its active size, marking a significant step towards more resource-efficient large language models. The release makes Hy3 accessible to developers and researchers globally.

Performance and Hallucination Reduction

A key claim from Tencent revolves around Hy3’s ability to maintain strong performance while addressing common challenges in large language models, particularly hallucination. In a blind evaluation involving 270 experts, Hy3 achieved a score of 2.67 out of 4, surpassing GLM-5.1, which scored 2.51.

Internal testing conducted by Tencent also indicated a substantial reduction in the model’s hallucination rate. The company reported a drop from 12.5 percent to 5.4 percent, effectively halving the incidence of generating factually incorrect or misleading information. This improvement is critical for applications requiring high reliability and factual accuracy.

Accessibility and Integration

Hy3 is now available under an Apache 2.0 license, promoting broad adoption and collaborative development within the AI community. The model can be accessed through popular platforms such as Hugging Face, ModelScope, and GitHub. For developers seeking further optimization, an FP8-quantized version is also provided.

Tencent has outlined plans to extend support for Hy3 to additional platforms, including OpenRouter and Cline, to broaden its reach. The company has already integrated Hy3 into several of its own products, demonstrating its practical applications. These include WorkBuddy, the Yuanbao chatbot, WeChat, and the game assistant for “Path of Exile: Advent.”

Key facts

Feature	Detail
Total Parameters	295 billion
Active Parameters	21 billion (plus 3.8 billion for MTP layer)
Architecture	Mixture-of-Experts (MoE)
Context Length	Up to 256,000 tokens
Hallucination Rate	Reduced from 12.5% to 5.4%
Licensing	Apache 2.0

Implications for Indian AI Development

For the Indian AI ecosystem, the release of Hy3 offers several potential benefits. The open-source nature of the model, combined with its claimed efficiency and reduced hallucination rates, could provide a valuable resource for Indian startups and researchers. Developing high-performing AI applications often requires significant computational resources, which open-source models like Hy3 can help mitigate by offering competitive performance at potentially lower operational costs.

The availability of an FP8-quantized version is particularly relevant, as it allows for deployment on hardware with more constrained resources, making advanced AI capabilities more accessible across various applications in India, from digital services to enterprise solutions.

Source: The Decoder, “Tencent releases Hy3 open-source model that allegedly matches models up to five times its active size”