tencent/Hy3 New Apache 2.0 licensed model from Tencent in China: Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team. Following the Hy3 Preview launch in late April, we gathered feedback from 50+ products and scaled up post-training with higher quality data. Today, we introduce Hy3, which outperforms similar-size models and rivals flagship open-source models with 2-5x parameters. It also shows signifi ✦ AI generated
The full Hy3 release builds on feedback from over 50 products that used the earlier Preview version, incorporating improved post-training data.
Tencent Hy Team: Following the Hy3 Preview launch in late April, we gathered feedback from 50+ products and scaled up post-training with higher quality data.
Tencent describes the technical architecture of its new Hy3 model: a 295B-parameter MoE model with 21B active parameters.
Tencent Hy Team: Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team.
Simon Willison notes Hy3's temporary free access on OpenRouter and tried his signature pelican-on-a-bicycle SVG benchmark on it.
Simon Willison: It's available for free on OpenRouter until July 21st. I had it "Generate an SVG of a pelican riding a bicycle" there and got this:
Tencent claims Hy3 beats models of comparable size and is competitive with much larger flagship open-source models.
Tencent Hy Team: Today, we introduce Hy3, which outperforms similar-size models and rivals flagship open-source models with 2-5x parameters.
explains mechanism · 1
Beyond benchmark performance, Tencent says Hy3 delivers measurable practical improvements in real product and productivity use cases.
Tencent Hy Team: It also shows significant gains in utility across various products and productivity tasks.
explains mechanism · 1
Hy3 is a large download even quantized, but supports a long 256K-token context window.
Simon Willison: The full-sized model is 598GB on Hugging Face, and the FP8 quantized one is 300GB. The context length is 256K.