Tencent Releases Hy3, a Powerful New Open-Source AI Model
Tencent has launched Hy3, a new open-source AI model licensed under Apache 2.0, developed by the Tencent Hy Team. This Mixture-of-Experts (MoE) model boasts 295 billion parameters, with 21 billion active parameters and 3.8 billion parameters in its MTP layers. Following a preview launch in late April and subsequent feedback from over 50 products, Hy3 has undergone post-training with enhanced data quality. The model demonstrates superior performance compared to similarly sized models and competes with leading open-source models that have two to five times more parameters. It also shows marked improvements in utility for various products and productivity tasks. The full-sized Hy3 model occupies 598GB on Hugging Face, while an FP8 quantized version is available at 300GB. It supports a context length of 256K tokens. Hy3 is accessible for free on OpenRouter until July 21st.
The release of Tencent's Hy3 model signifies a continued acceleration in the development and accessibility of large language models, particularly from non-Western entities. By adopting an open-source license, Tencent not only fosters broader adoption and community-driven innovation but also strategically positions itself within the global AI ecosystem. The model's performance claims, rivaling larger proprietary models, suggest a maturing efficiency in parameter utilization and training methodologies. This development highlights the increasing competitiveness of AI research originating from China and its potential to influence global AI standards and applications. The availability of a free tier on OpenRouter until July 21st serves as a direct incentive for developers and businesses to experiment with and integrate Hy3, potentially leading to novel applications and further insights into its capabilities and limitations.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.