Tencent's open-weight Hy3 model surges up Hugging Face's trending charts

Tencent’s Hy3, an open-weight language model released by its Hunyuan team on July 6, 2026, surged up Hugging Face’s trending charts on July 9, drawing a renewed wave of developer interest three days after launch.

The renewed attention centers on an efficiency claim. Hy3 is a 295-billion-parameter Mixture-of-Experts (MoE) system that activates only 21 billion parameters per token, and Tencent says it delivers intelligence comparable to flagship models with two to five times its parameter count.

Released under an Apache 2.0 license to Hugging Face, ModelScope and other repositories, Hy3 succeeds the Hy3 Preview from April 23, according to Tencent’s release post. The company says the model cut hallucination rates from 12.5% to 5.4% and that average daily token consumption has risen twentyfold since the preview.

Reported benchmark scores on the model card include 78.0 on SWE-Bench Verified and 90.4 on GPQA Diamond. A blind expert evaluation across 270 expert assessments scored Hy3 at 2.67 out of 4.0, against 2.51 for GLM-5.1. Those figures are Tencent’s own and have not been independently verified.

Hy3 supports a 256K-token context window and ships with an FP8 quantized variant, and Tencent says it is being integrated into third-party tools including OpenRouter, Cline and Cherry Studio, alongside its own WorkBuddy, CodeBuddy and Yuanbao products. The model is available through Tencent Cloud’s TokenHub API.