model_release
Tencent Releases Hy4 Preview with 770 Billion Parameters
Tencent has introduced Hy4 preview, a large Mixture-of-Experts AI model with 770 billion parameters and a context window of 1 million tokens. The model's weights are publicly available under the Apache 2.0 license, accessible via API and self-deployment platforms.
AS1 News
On August 28, Tencent announced Hy4 preview, a new addition to its Hy/Hunyuan series of AI models. This Mixture-of-Experts (MoE) model features 256 routed experts and 49 billion active parameters. The model's weights are available on Hugging Face, enabling developers to utilize it independently without API restrictions. Tencent also revealed pricing details: $0.834 per 1 million input tokens and $2.501 per 1 million output tokens, exerting competitive pressure on the market. Hy4 is tailored for tasks such as coding, document analysis, financial calculations, and game development.
A notable aspect of Hy4 is its participation in self-optimization processes—used to automatically enhance training methods, data strategies, and infrastructure—resulting in a 31.8% increase in system throughput. Internal benchmarks indicate a performance score of approximately 3 out of 4 based on expert evaluation, underscoring the model's high quality.
Tencent plans to release new versions of Hy4 in the near future, though specific dates have not been announced. The model is already accessible through multiple platforms, and its API pricing makes it a competitive option in the large language model market.
The release of Hy4 preview introduces a highly capable, open-weight AI model that could influence the development and deployment of large language models across various industries.