Tencent has released Hy3, a 295B-parameter Mixture-of-Experts model that rivals flagship open-source models while requiring significantly fewer active parameters. The model features a 256K context window and demonstrates improved performance in utility and productivity tasks following extensive post-training. It is available on Hugging Face and was offered for free access via OpenRouter in early July 2026.
Background
Tencent's Hy team continues to advance large language model capabilities in China, focusing on efficient architectures like Mixture-of-Experts to maximize performance per parameter. This release follows the Hy3 Preview launched in April 2026, incorporating feedback from over 50 products to enhance quality and utility.
- Source
- Simon Willison
- Published
- Jul 7, 2026 at 07:57 AM
- Score
- 7.0 / 10