Tencent has released Hy3, an open-source language model designed to deliver outsized performance from a compact footprint. The 295-billion-parameter model uses a mixture-of-experts architecture that activates only 21 billion parameters during inference, keeping computational costs low while maintaining competitive output quality.
The company claims Hy3 matches the performance of models two to five times larger than its active parameter count. This efficiency gain matters for deployment at scale. Running smaller active models reduces latency, power consumption, and hardware requirements across data centers. For enterprises, fewer active parameters translate directly to lower inference costs.
Tencent also reports a hallucination rate of 5.4 percent, half the rate of comparable models. Hallucinations, where language models generate false or nonsensical information presented as fact, remain a persistent problem in production systems. Cutting this rate significantly improves reliability for applications where accuracy is non-negotiable, such as customer support automation or medical information systems.
The mixture-of-experts approach has gained traction across the industry. Instead of activating all parameters for every input token, MoE models route different types of queries to specialized sub-networks. This selective activation lets developers pack more total capacity into a model while keeping per-inference computation manageable. Google's Gemini and Meta's earlier research both explored this approach.
Open-sourcing Hy3 signals Tencent's bet on community adoption and external validation. Public models face real-world testing that proprietary systems cannot match. Third-party benchmarking will quickly reveal whether Hy3's performance claims hold across diverse tasks beyond Tencent's testing conditions.
The release positions Tencent as a serious contender in the open-source language model space, competing directly with Meta's Llama variants and Mistral's offerings. As enterprises