Tencent's Hy Team has released Hy3, a Mixture-of-Experts (MoE) large language model. The model contains 295 billion parameters in total, including 21 billion active parameters and 3.8 billion parameters in the MTP layer.
Development and refinement
A Hy3 preview was published in late April. After that initial release, the team collected feedback from over 50 product teams and performed post-training scaling using higher-quality data. Tencent says these iterations improved the model's overall performance and practical usefulness.
Performance and use cases
According to Tencent, Hy3 outperforms models of comparable size and competes with flagship open-source models that have two to five times more parameters. The company also reports notable utility gains across various products and productivity-oriented tasks.
Size, formats and availability
The full-precision model occupies 598 GB on Hugging Face, while the FP8-quantized version is about 300 GB. Hy3 supports a context length of 256,000 tokens (256K).
The model is available for free through OpenRouter until July 21.
Demonstration example
As a simple demonstration, the announcement notes that Hy3 successfully generated an SVG of a pelican riding a bicycle when prompted, illustrating one facet of its generative capabilities.
Why this matters
Hy3's launch underscores the continued interest in MoE architectures for scaling large models: MoE allows very large parameter counts while activating only portions of the network per request, which can reduce computation cost relative to dense models. The long 256K context window is especially relevant for handling long documents, codebases, or extended multi-turn conversations.
According to Tencent Hy Team, Hy3 is intended to deliver stronger performance and increased practical value across a range of products and development scenarios.
