Chinese technology giant Tencent has released a new open-source artificial intelligence model, Hy3, claiming it can match leading AI systems while using far fewer parameters than many of its rivals.

Launched on July 6, Hy3 is Tencent’s latest reasoning and AI agent model built on a Mixture-of-Experts (MoE) architecture. The model contains 295 billion total parameters, but activates only 21 billion during inference, making it significantly more efficient than several competing open-source models.

According to Tencent’s published benchmark results, Hy3 delivers performance comparable to major open models including GLM-5.2 and DeepSeek-V4-Pro. The company also reported that Hy3 outperformed GPT-5.5 on the FrontierScience-Olympiad benchmark, which measures an AI model’s ability to solve advanced scientific reasoning and research tasks.

Despite being considerably smaller than its competitors—GLM-5.2 has 753 billion parameters while DeepSeek-V4-Pro has 1.6 trillion—Tencent says Hy3 achieves similar performance with substantially lower computational requirements.

Tencent said the model was trained with strict data-quality controls designed to reduce hallucinations, factual errors and logical inconsistencies. The company added that Hy3 is designed to provide answers only when supported by evidence and to explicitly acknowledge uncertainty when sufficient information is unavailable.

The company also cited internal testing of its AI assistant WorkBuddy, claiming Hy3 completed document-processing tasks using 47.4% fewer tokens than GLM-5.2, resulting in lower computing costs and improved efficiency.

Tencent has released Hy3 as an open-source model under the Apache 2.0 licence, allowing developers and businesses to use, modify and deploy the model for both commercial and research purposes.