tencent/Hy3

This article covers the release of Hy3, a 295B-parameter Mixture-of-Experts (MoE) AI model by Tencent's Hy Team, with 21B active parameters and 256K context length. It outperforms similar-size models and rivals flagship open-source models with 2-5x parameters. The model is Apache 2.0 licensed, available in full (598GB) and FP8 quantized (300GB) versions on Hugging Face, and free on OpenRouter until July 21st. The author tests its SVG generation capability, comparing it to an earlier preview.