release
tencent/Hy3
Hy3 offers developers a powerful, open-source model with competitive performance, accessible for free testing via OpenRouter, enabling cost-effective experimentation in AI workflows.
What happened
Tencent has released Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) language model under the Apache 2.0 license, as reported by Simon Willison. The model has 21 billion active parameters and a 3.8-billion-parameter MTP layer, making it efficient while maintaining high capacity. Hy3 outperforms models of comparable size and competes with larger open-source models up to five times its parameter count, according to Tencent. The full model is available on Hugging Face (598 GB), with an FP8 quantized version (300 GB) and a context length of 256K tokens. For developers, Hy3 is freely accessible via OpenRouter until July 21, offering a low-risk way to evaluate its capabilities. Willison tested it by generating an SVG of a pelican on a bicycle. This release highlights the growing competitiveness of Chinese AI labs in open-source large language models, providing builders with another powerful, cost-effective option for AI workflows.
Key takeaways
- Tencent released Hy3, a 295B-parameter MoE model with 21B active parameters, under Apache 2.0.
- Model outperforms similar-size models and rivals flagship open-source models with 2-5x parameters.
- Available on Hugging Face in full (598 GB) and FP8 quantized (300 GB) versions; context length 256K.
- Free access on OpenRouter until July 21, enabling easy testing for developers.
- Source: Simon Willison, who generated an SVG of a pelican riding a bicycle using the model.
Why it matters
Hy3 offers developers a powerful, open-source model with competitive performance, accessible for free testing via OpenRouter, enabling cost-effective experimentation in AI workflows.
This is an original editorial digest by AI Workflow Center. Full reporting at the source:
Read the original on Simon WillisonMore AI news
All news →
Run Your Own AI Directory