Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
Data Science
Simon Willison's Blog·Jul 6·5 min read
tencent/Hy3
TL;DR
Tencent has launched Hy3, a 295B-parameter Mixture-of-Experts model that utilizes 21B active parameters. It significantly outperforms similar models and rivals leading open-source options.
✦ Why It Matters
Engineers can experiment with Hy3 on OpenRouter to enhance their applications with advanced AI capabilities.
Key Takeaways
- Test Hy3 for specific productivity tasks in your applications.
- Consider the benefits of using a Mixture-of-Experts architecture.
- Evaluate the performance of the FP8 quantized model for efficiency.
- Provide feedback to Tencent if using Hy3 in your projects.
Full Summary
Hy3 is a new language model developed by Tencent's Hy Team, featuring a total of 295 billion parameters, of which 21 billion are active during operation. This Mixture-of-Experts (MoE) architecture allows for efficient scaling, utilizing only a subset of parameters at any time.
Following a preview in April and feedback from over 50 products, Hy3 has been optimized with higher quality data. The model is available in two sizes: a full version at 598GB and a quantized version at 300GB, with a context length of 256K tokens.
Initial tests indicate that Hy3 outperforms similar-sized models by 2-5 times in various productivity tasks, making it a strong contender in the AI landscape. It is currently accessible for free on OpenRouter until July 21st.